Mapping and Exploration for Search and Rescue with Humans and ...

Dissertation zur Erlangung des Doktorgrades der Fakultät für Angewandte
Wissenschaften der Albert-Ludwigs-Universität Freiburg im Breisgau
Mapping and Exploration for Search and
Rescue with Humans and Mobile Robots
Alexander D. Kleiner
2007

2
Tag der Disputation:
20. Februar 2008
Dekan:
Prof. Dr. Bernhard Nebel, Albert-Ludwigs-Universität Freiburg
Gutachter:
Prof. Dr. Bernhard Nebel, Albert-Ludwigs-Universität Freiburg
Prof. Dr. Daniele Nardi, University of Rome “Sapienza”

Zusammenfassung
Katastrophenhilfe ist eine zeitkritische Aufgabe bei der Überlebende innerhalb der ersten 72
Stunden gerettet werden müssen. Ein Ziel der Rettungsrobotik ist die Unterstützung dieser
Aufgabe mit kooperierenden Teams bestehend aus Menschen und Robotern. Dabei sollen das
Katastrophengebiet flächendeckend erkundet und Positionen Überlebender an eine zentrale Befehlsstelle
zur Planung von Rettungsmissionen gemeldet werden. Dies kann nur effizient erfolgen
wenn Menschen und Roboter das Katastrophengebiet gemeinsam kartieren und gleichzeitig
ihre Suche koordinieren. Roboter sollten dabei in der Lage sein, Teilprobleme, wie z.B. Navigation
und das Aufspüren von Opfern, autonom durchzuführen. Aufgrund der unstrukturierten
Beschaffenheit von Katastrophengebieten und der Unzugänglichkeit des Geländes ist die Bewältigung
dieser Aufgaben außerordentlich schwierig. Des Weiteren müssen entwickelte Lösungen
auch bei einem Totalausfall der Kommunikation dezentral funktionsfähig sein.
In dieser Dissertation wird ein vereinheitlichter Ansatz für Menschen und Roboter zur Lösung
dieser Probleme vorgestellt. Dabei werden Kernprobleme, wie z.B. Positionsbestimmung
auf unwegsamen Gelände, Kartenerstellung durch heterogene Teams und dezentrale Team Koordination
mit eingeschränkter Kommunikation direkt behandelt. Es wird die Methode “RFID-
SLAM” zur robusten und effiziente Kartierung großräumiger Gebiete vorgestellt, welche RFID
Technologie zur Wiedererkennung von Orten einsetzt. Die vorgestellte Methode beinhaltet
einerseits die fortwährende Positionsbestimmung von Menschen und Robotern, jeweils mittels
Fußgänger-Koppelnavigation und rutschemfindlicher Radodometrie, und andererseits eine
graphenbasierte Kartenoptimierung. Das Verfahren kann zentral oder dezentral erfolgen, wobei
im Gegensatz zu herkömmlichen Methoden, ein wiederholtes Besuchen von Orten einzelner
Teammitglieder zum optimieren der Karte nicht erforderlich ist. Aufbauend auf dieser Kartenrepräsentation
werden für die Koordination größerer Rettungsteams eine deliberative Methode
zur Bestimmung der Aufgabenverteilung und Multiagenten-Pfadplanung und eine Methode zur
lokalen Suche unter der Verwendung des Speichers von RFIDs, vorgestellt.
Zur autonomen Navigation von Robotern auf unwegsamen Gelände und automatischen Erkennung
von Menschen in Katastrophengebieten werden einerseits eine effiziente Methode zur Erstellung
von Höhenkarten und andererseits ein Ansatz zur genetischen Optimierung von “Markov
Random Field” Modellen präsentiert. Abschließend wird eine human in the loop Architektur
vorgestellt, welche die Integration von gesammelten Daten einzelner Rettungseinheiten in ein
Multiagentensystem erlaubt. In diesem Zusammenhang wird die zentrale Koordination großräumiger
Katastrophenhilfe mittels Agententechnologie beschrieben.
Die in dieser Arbeit vorgestellten Verfahren wurden weitgehend in Outdoor-Szenarien und offiziellen
Testszenarien zum Katastrophenschutz getestet. Sie waren ein wesentlicher Bestandteil
von Systemen, die insgesamt mehr als zehn mal den ersten Platz bei internationalen Wettbewerben,
wie z.B. den RoboCup Weltmeisterschaften, erreichten.
i

Abstract
Urban Search And Rescue (USAR) is a time critical task since all survivors have to be rescued
within the first 72 hours. One goal in Rescue Robotics is to support emergency response by
mixed-initiative teams consisting of humans and robots. Their task is to explore the disaster
area rapidly while reporting victim locations and hazardous areas to a central station, which
then can be utilized for planning rescue missions.
To fulfill this task efficiently, humans and robots have to map disaster areas jointly while coordinating
their search at the same time. Additionally, robots have to perform subproblems, such
as victim detection and navigation, autonomously. In disaster areas these problems are extraordinarily
challenging due to the unstructured environment and rough terrain. Furthermore, when
communication fails, methods that are deployed under such conditions have to be decentralized,
i.e. operational without a central station.
In this thesis a unified approach joining human and robot resources for solving these problems
is contributed. Following the vision of combined multi-robot and multi-human teamwork, core
problems, such as position tracking on rough terrain, mapping by mixed teams, and decentralized
team coordination with limited radio communication, are directly addressed. More specific,
RFID-SLAM, a novel method for robust and efficient loop closure in large-scale environments
that utilizes RFID technology for data association, is contributed. The method is capable of
jointly improving multiple maps from humans and robots in a centralized and decentralized
manner without requiring team members to perform loops on their routes. Thereby positions
of humans are tracked by PDR (Pedestrian Dead Reckoning), and robot positions by slippagesensitive
odometry, respectively. The joint-graph emerging from these trajectories serves as an
input for an iterative map optimization procedure. The introduced map representation is further
utilized for solving the centralized and decentralized coordination of large rescue teams. On
the one hand, a deliberate method for combined task assignment and multi-agent path planning,
and on the other hand, a local search method using the memory of RFIDs for coordination, are
proposed.
For autonomous robot navigation on rough terrain and real-time victim detection in disaster
areas an efficient method for elevation map building and a novel approach to genetic MRF
(Markov Random Field) model optimization are contributed. Finally, a human in the loop architecture
is presented that integrates data collected by first responders into a multi-agent system
via wearable computing. In this context, the support and coordination of disaster mitigation in
large-scale environments from a central-command-post-perspective are described.
Methods introduced in this thesis were extensively evaluated in outdoor environments and
official USAR testing arenas designed by the National Institute of Standards and Technology
(NIST). Furthermore, they were an integral part of systems that won in total more than 10 times
the first prize at international competitions, such as the RoboCup world championships.
i

To Wei Fang ( ) and my family,
for their love and support,
who provided the foundation
of all my endeavors.

Acknowledgements
My work was accompanied and supported by many people which took influence by
different means. In my opinion, there are basically three different categories of good
influence: motivation/virtue, team work/learning experience, and entertainment, which
are all equally important. Although these three categories overlap with either friends or
colleagues, and in some cases influence came through simply exchanging a few words,
and in other cases through years of joint work, I would like to thank them in this way.
First of all, I would like to thank my adviser Professor Bernhard Nebel who generously
made the building of each robot and agent system that we designed financially
possible, but more importantly, who motivated me. Furthermore, his integrity and correctness
polarized my personal development from the very beginning, and I very much
appreciate the liberties I received to work on my projects.
During my first years in Freiburg I very much enjoyed the time with the CS Freiburg
robot soccer team, particularly the work with Thilo Weigel, Markus Dietl, and Steffen
Gutmann. From the nights I spent with them in the laboratory or at RoboCup venues I
learned, above all, endurance and determination, which I later carried on to many other
teams.
During visits of Rescue Robotics related conferences and events I met a lot of highly
active people in the community, which greatly motivated my work. Hence, I would
like to thank Adam Jacoff for his restless efforts in setting up every year new performance
metrics, even if it means to saw countless cubic meters of wood. Furthermore,
he motivated robotics researchers by enabling discussions on disaster mitigation with
emergency responders from the field. My work was also motivated by the commitment
of Prof. Satoshi Tadokoro, Prof. Mike Lewis, and Prof. Daniele Nardi. I like to thank
them for all the efforts they have undertaken for the community, and also for giving
research directions.
Many thanks to Dirk Hähnel for the straightforward discussions we had on SLAM,
and supporting our team while building-up the first rescue robot.
Many thanks to my friends and colleagues for the wonderful team work we experienced.
First of all I would like to thank Christian Dornhege, who accompanied me for
the longest time. Besides his broad knowledge and ability to solve problems practically,
his diligence and discipline enabled many systems to function as they should. Furthermore,
I like to thank Rainer Kümmerle and Bastian Steder for their brilliant work with
Rescue Robots Freiburg. It is their individual skills that lay the basis for strong team
work, particularly Basti’s ability to solve any mechanical and electronic problem by
iii

iv
Acknowledgements
a hot-glue gun, cable fixer and shrink tubing, and Rainer’s understanding of putting
mathematics into source code. Thanks to all other team members contributing essential
parts to the systems that we built with ResQ Freiburg and Rescue Robots Freiburg
from 2004-2006: Tobias Bräuer, Michael Brenner, Jörg Stückler, Moritz Göbelbecker,
Mathias Luber, Daniel Meyer-Delius, Johann Prediger, Michael Ruhnke, and Michael
Schnell. Also many thanks to Dali Sun who walked several kilometers to collect PDR
(Pedestrian Dead Reckoning) data evaluated in this thesis.
Vittorio Amos Ziparo (Don Vito) from Italy visited our laboratory for one year as a
guest researcher. On the one hand his way of making jokes and taking things easy made
working together very pleasant. On the other hand his strong motivation to bring things
to an end, which I was certain about as he still did not want to sleep after two days and
nights of intensive work, and his way of taking the consequences from decisions made,
finally lead to the success we had.
I would like to thank Holger Kenn, for the nice team work we had, his optimistic
thinking, and his directness. Anyway, he was the first person working on rescue robots
who I met.
My thanks to all my friends and colleagues for the great time I had in Freiburg. Special
thanks to Michael Brenner for all the meaningful discussions on multi-agent planning
and also for enduring me in our office for so long. Thanks to Sebastian Kupferschmidt,
Dapeng Zhang, and Malte Helmert for having table sports and beer garden
events. Also many thanks to Gian Diego Tipaldi and Giorgio Grisetti for the late night
excursions, as well as for teaching me on SLAM.
Finally, most thanks to my oldest friend Jörg Ziegler, my family Julia, Brigitte, and
Dieter Kleiner, and most of all Wei Fang, for supporting me and enjoying life with me.

Table of Contents
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 RoboCup Rescue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Structure of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5 Relevant publications and awards . . . . . . . . . . . . . . . . . . . . . 10
2 Rescue Robotics Hardware 15
2.1 Design criteria for robot platforms . . . . . . . . . . . . . . . . . . . . 15
2.2 Sensors for navigation . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.1 Inertial Measurement Unit (IMU) . . . . . . . . . . . . . . . . 18
2.2.2 Laser Range Finder (LRF) . . . . . . . . . . . . . . . . . . . . 20
2.2.3 Global Navigation Satellite System (GNSS) . . . . . . . . . . . 21
2.3 Sensors for detecting humans . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.1 Video and infra-red cameras . . . . . . . . . . . . . . . . . . . 24
2.3.2 3D Laser Range Finder (LRF) . . . . . . . . . . . . . . . . . . 25
2.3.3 Audio and CO 2 sensors . . . . . . . . . . . . . . . . . . . . . . 26
2.4 RFID technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.1 RFIDs in robotic applications . . . . . . . . . . . . . . . . . . 30
2.4.2 Importance for Urban Search and Rescue (USAR) . . . . . . . 30
2.5 Rescue Robots Freiburg . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5.1 Lurker robot . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5.2 Zerg robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5.3 Human-Robot Interaction (HRI) . . . . . . . . . . . . . . . . . 34
3 Pose Tracking In Disaster Areas 37
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
v

vi
Table of Contents
3.2 Conventional techniques . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2.1 Dead reckoning . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2.2 Scan matching . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3 Slippage sensitive wheel odometry . . . . . . . . . . . . . . . . . . . . 42
3.4 Visual odometry based pose tracking . . . . . . . . . . . . . . . . . . . 43
3.4.1 Feature tracking . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4.2 Filtering of rotations around the pitch angle . . . . . . . . . . . 45
3.4.3 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.5 Pedestrian Dead Reckoning (PDR) . . . . . . . . . . . . . . . . . . . . 49
3.5.1 Distance estimation . . . . . . . . . . . . . . . . . . . . . . . . 49
3.5.2 Heading estimation . . . . . . . . . . . . . . . . . . . . . . . . 51
3.6 Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.6.1 Pose tracking under heavy slippage . . . . . . . . . . . . . . . 52
3.6.2 Pose tracking on tracked vehicles . . . . . . . . . . . . . . . . 54
3.6.3 Pose tracking of human walking . . . . . . . . . . . . . . . . . 59
3.7 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4 RFID Technology-based SLAM 63
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.2 Conventional techniques . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2.1 EKF-based SLAM . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2.2 FastSLAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.3 Centralized RFID-SLAM . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.3.1 Building RFID graphs . . . . . . . . . . . . . . . . . . . . . . 69
4.3.2 Linear optimization . . . . . . . . . . . . . . . . . . . . . . . . 70
4.3.3 Non-linear optimization . . . . . . . . . . . . . . . . . . . . . 72
4.3.4 Trajectory interpolation . . . . . . . . . . . . . . . . . . . . . . 74
4.3.5 Sensor model . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.4 Decentralized RFID-SLAM (DRFID-SLAM) . . . . . . . . . . . . . . 75
4.5 Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.5.1 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . 78
4.5.2 RFID-SLAM with robots . . . . . . . . . . . . . . . . . . . . . 80
4.5.3 RFID-SLAM with humans . . . . . . . . . . . . . . . . . . . . 85

Table of Contents
vii
4.5.4 RFID-SLAM jointly with humans and robot . . . . . . . . . . . 88
4.6 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5 Real-time Elevation Mapping 93
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.2 Conventional techniques . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.2.1 Occupancy grid maps . . . . . . . . . . . . . . . . . . . . . . . 94
5.3 Building elevation maps in real-time . . . . . . . . . . . . . . . . . . . 96
5.3.1 Single cell update from sensor readings . . . . . . . . . . . . . 96
5.3.2 Map filtering with a convolution kernel . . . . . . . . . . . . . 99
5.3.3 Estimation of the three-dimensional pose . . . . . . . . . . . . 100
5.4 Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.5 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6 RFID Technology-based Multi-Robot Exploration 107
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.2 Coordinated local exploration . . . . . . . . . . . . . . . . . . . . . . . 109
6.2.1 Navigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.2.2 Local exploration . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.3 Multi-robot topological maps . . . . . . . . . . . . . . . . . . . . . . . 112
6.4 Global exploration monitoring . . . . . . . . . . . . . . . . . . . . . . 114
6.4.1 Problem modeling . . . . . . . . . . . . . . . . . . . . . . . . 114
6.4.2 Global task assignment and path planning . . . . . . . . . . . . 117
6.4.3 Monitoring agent . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.5.1 Evaluation of the local approach . . . . . . . . . . . . . . . . . 120
6.5.2 Evaluation of the global approach . . . . . . . . . . . . . . . . 122
6.6 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
7 Victim Detection 129
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.2 Vision data processing . . . . . . . . . . . . . . . . . . . . . . . . . . 130

viii
Table of Contents
7.3 Genetic model optimization . . . . . . . . . . . . . . . . . . . . . . . . 132
7.3.1 Markov Random Fields (MRFs) . . . . . . . . . . . . . . . . . 132
7.3.2 Model selection . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
7.5 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
7.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
8 Human In The Loop 143
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
8.2 Interfacing human first responders . . . . . . . . . . . . . . . . . . . . 144
8.2.1 Preliminary test system . . . . . . . . . . . . . . . . . . . . . . 145
8.2.2 Wearable emergency-response system . . . . . . . . . . . . . . 146
8.3 Multi Agent Systems (MAS) for disaster management . . . . . . . . . . 147
8.3.1 Real-world data integration . . . . . . . . . . . . . . . . . . . . 148
8.3.2 Coordinated victim search . . . . . . . . . . . . . . . . . . . . 149
8.3.3 Rescue sequence optimization . . . . . . . . . . . . . . . . . . 153
8.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
8.4.1 Data integration from wearable devices . . . . . . . . . . . . . 155
8.4.2 Coordinated victim search . . . . . . . . . . . . . . . . . . . . 155
8.4.3 Victim rescue sequence optimization . . . . . . . . . . . . . . . 158
8.5 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
8.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
9 Summary And Future Work 163
9.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
9.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
10 Appendix 169
10.1 Rescue communication protocol . . . . . . . . . . . . . . . . . . . . . 169
10.2 Construction drawings . . . . . . . . . . . . . . . . . . . . . . . . . . 173
10.2.1 3D Laser Range Finder (LRF) . . . . . . . . . . . . . . . . . . 173
10.2.2 Zerg robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
10.2.3 RFID tag releaser . . . . . . . . . . . . . . . . . . . . . . . . . 179
10.3 Schematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194

1 Introduction
1.1 Motivation
Large scale urban disaster situations, such as earthquakes, floods or terrorist attacks
pose an immense threat to modern urban civilization centers. Examples such as the
Great Hanshin earthquake in Kobe, the Pacific Tsunami, the hurricane Kathrina or the
terrorist attacks of 9/11 demonstrate that although disaster response plans were in place,
these events pushed existing countermeasures over their limits, resulting in loss of human
lives, and long-term adversary effects on the affected regions.
During Urban Search And Rescue (USAR), time is of utmost importance. Rescue
teams have to explore large terrain within a short amount of time in order to locate survivors
after a disaster. In this scenario, the number of survivors decreases drastically by
each day due to hostile environmental conditions and the victims’ injuries. Therefore,
the survival rate depends significantly on the efficiency of rescue teams.
Furthermore, cooperation is a must during rescue operations for disaster mitigation.
In general the problem is not solvable by a single unit, yet a heterogeneous team that dynamically
combines individual capabilities in order to solve the task is needed [Murphy
et al., 2000]. This is due to the structural diversity of disaster areas, variety of evidence
on human presence that sensors can perceive, and the necessity to quickly and reliably
examining targeted regions. Mixed teams not only offer the possibility to field such diverse
capabilities, they also exhibit increased robustness due to redundancy [Dias et al.,
2004], and increased performance due to parallel task execution.
Particularly the aspect of integrating mobile robots in human search teams, and thus
to combine robotics technology with human expertise, attracted increasing attention
in recent years. One goal there is to build mixed-initiative teams [Wang and Lewis,
2007], consisting of autonomous robots that are partially controlled by human incident
commanders. Their task is to jointly create a map of the disaster area and to register
victim locations, which are further utilized by human task forces to schedule rescue
missions.
To deploy robots for exploration and mapping has multiple advantages. Miniature
robots can enter either confined or toxic spaces that humans and search dogs cannot.
Search dogs are of great assistance during emergency response. However, at the WTC
at 9/11 they were hindered by rain, which limited their sense of smell to only 0.3
meters. While victims were assumed to be located at a depth of 30 meters they could
1

2 Chapter 1. Introduction
(a)
(b)
Figure 1.1: First responders and destroyed building structures after the 9/11 incident.
Courtesy of Robin Murphy.
not go deeper than half a meter into the rubble because there was no air [AAAI, 2002].
Furthermore, rescue safety and effectiveness is a serious issue. In the case of the Mexico
City earthquake in 1985, 135 rescuers died, many of them while searching confined
spaces that were flooded [Casper et al., 2000]. Figure 1.1 depicts collapsed building
structures after 9/11.
Rescue workers have not enough time to thoroughly search the area since they are
needed for rescuing as many victims as possible within the first 72 hours, which are
therefore also called the golden 72 hours. After this time, the survival rate decreases
drastically (see Figure 1.2). For example, it takes in average 10 trained professionals approximately
4 hours to remove a victim in a void space and 10 trained professionals 10
hours to remove an entombed victim [Administration and Agency, 1995]. Here, robotic
technology can be allocated to find victims thereby enabling the humans to concentrate
on the rescue. Also, robotic technology can influence rescue missions substantially by
providing information on human activities in the field. Tom Haus, a firefighter at 9/11,
pointed out these problems when I met him at the NIST emergency responder exercise.
He depicted the difficulties first responders encounter in terms of situation awareness
and keeping orientation. He noticed during a panel discussion: “We need a tracking
system that tells us where we are, where we have been, and where we have to go to”.
A thorough analysis of the requirements for the performance of robots that can
support USAR roles and tasks has been undertaken by the US authorities NIST (National
Institute of Standards and Technology) and DHS (Department of Homeland Security)
[Messina et al., 2005]. They acquired expert knowledge by polling members
from twenty FEMA (Federal Emergency Management Agency) task forces and the US
National Guard. Among other USAR tasks, they identified reconnaissance and primary
search as the two highest priorities for applying robots.

1.1. Motivation 3
Number
1600
1400
1200
1000
800
600
400
200
0
16 18 20 22 24 26 28 30 32
Date (Jan. 1995)
(a)
Dead
Survive
Rescued total
Survival Rate [%]
100
90
80
70
60
50
40
30
20
10
0
30 minutes
1 Day
2 Day 3 Day
Time
(b)
4 Day
5 Day
Figure 1.2: Disaster mitigation statistics: (a) death and survival after the Kobe earthquake,
(b) survival rate from the Federal Emergency Management Agency (FEMA).
Courtesy of Satoshi Tadokoro.
Due to the fact that Rescue Robotics faces extraordinary challenging problems, such
as unstructured environments and hostile conditions, qualitative and quantitative performance
metrics are indispensable. During the last years, researchers and organizations
developed such performance metrics in order to asses the deployment of robots for disaster
mitigation. Casper and Murphy analyzed experiences from a robot deployment
at 9/11 [Casper and Murphy, 2003]. They concluded that robot portability and Human
Robot Interfaces (HRI) are the most pressing needs. Burke et al. examined operator situation
awareness during tele-operated search team interaction during a disaster rescue
drill by analyzing communications. Pratt et al. photo documented with an iSENSYS IP3
micro air vehicle damage in multi-story buildings after hurricane Kathrina [Pratt et al.,
2006]. They provide logged sensor data and useful hints on their homepage [CRASAR,
2007]. The special project for Earthquake Disaster Mitigation in Urban Areas (DDT
Project), started by the Japanese Ministry of Education, Sports, Culture, Science and
Technology in 2002, concentrates on the development of robotic solutions for reconnaissance
of victim bodies, structural damage and environmental conditions to assist
rescue teams in large-scale urban earthquake disasters [Tadokoro, 2005,Tadokoro et al.,
2003]. Furthermore, NIST developed a USAR test-bed for the RoboCup/AAAI robot
urban search and rescue competition with the goal to promote research in designing
intelligent fieldable robots and to provide a benchmark for testing current robot platforms
[Jacoff et al., 2000, Kitano and Tadokoro, 2001] (see Section 1.3).
One research challenge in Rescue Robotics is to acquire a map of the environment.
This involves autonomously solving in real-time the problem of Simultaneous Localization
and Mapping (SLAM), consisting of a continuous state estimation problem and
a discrete data association problem. In disaster areas, these problems are extraordinarily
challenging. On the one hand, state estimation is difficult due to the extremely unreli-

4 Chapter 1. Introduction
able odometry measurements usually found on robots operating on rough terrain. On
the other hand, data association, i.e. to recognize locations from sensor data, is challenging
due to the unstructured environment. Collapsed building structures, and limited
visibility conditions due to smoke, dust, and fire, prevent an easy distinction of different
places. This problem is also relevant to standard SLAM approaches, which recognize
places by associating vision-based and LRF-based features. These extraordinary circumstances
make it very difficult to apply common techniques from robotics. Many
of these techniques have been developed under strong assumptions, for example, they
require polygonal structures, such as typically found in office-like environments [Gutmann
and Schlegel, 1996, Grisetti et al., 2002] or depend on predictable covariance
bounds from pose tracking for solving the data association problem by validation gating
[Dissanayake et al., 2001]. Moreover, SLAM methods work with the principle
of map improvement through loop-closure, i.e. to improve the map globally each time
places have been re-observed. However, when facing the reality of emergency response,
firefighters will intentionally try to avoid performing loops, e.g. while they are searching
for victims.
To facilitate teamwork among humans and robots is another challenge in Rescue
Robotics. One difficulty that arises is that designing approaches for supporting emergency
response from a single robot/human perspective is prone to yield suboptimal
solutions. The joint performance of a team depends on assembling the right mixture
of individual human and robot capabilities, and thus has to be designed as a whole.
Another difficulty that arises is the exchange of information, such as map data and victim
locations, e.g. requiring reliable methods for merging individually generated map
pieces. While there has been extensive research on building maps from a single agent
perspective, and also on combining multiple maps offline, there has been only little
attention on addressing the problem of decentralized SLAM, i.e. to jointly improve a
map online based on sporadic point-to-point communications between the agents. In
emergency response scenarios, this issue is of major importance since communication
bandwidth is a truly limited resource. This also affects the problem of team coordination,
i.e. requiring approaches being operational even if communication is not possible
at all. Finally, robots have to be capable of detecting victims in the field, which is rather
difficult in real-time given arbitrary shapes and colors found in disaster areas.
1.2 Contribution
In this thesis a unified approach combining human and robot resources for supporting
disaster mitigation is presented. Whereas the main contribution is on autonomous mapping
and exploration of disaster areas, also solutions for human detection and top-level
coordination of rescue teams are presented. Following the vision of building an overall
approach of combined multi-robot and multi-human teamwork, core problems, such as

1.2. Contribution 5
position tracking on rough terrain, joint route graph optimization of multiple units, and
team coordination with limited radio communication are directly addressed. Conventional
approaches for solving these problems are scrutinized, and improved solutions
regarding efficiency and robustness in harsh environments are presented. The thesis
describes the following contributions, which will be further explained in detail:
• The design and construction of low-cost robotic hardware for search and rescue
research
• A method for slippage sensitive pose tracking on wheeled vehicles
• A method for visual odometry-based pose tracking on tracked vehicles
• RFID-SLAM: an algorithm for decentralized and centralized SLAM
• A local search method for decentralized team coordination
• A deliberative approach for centralized team coordination
• A method for building elevation maps on rough terrain in real-time
• A method for genetic optimization of MRF (Markov Random Field) models for
real-time victim detection
• A human-in-the-loop architecture building upon the RoboCup Rescue simulation
kernel for disaster mitigation
RFID-SLAM is a novel method for robust and efficient loop closure in large-scale
environments that utilizes RFID tags for data association. RFID tags are small devices
that carry a worldwide unique number that can be read from distant places, hence,
offering a robust way to label and recognize locations in unstructured environments.
Passive RFID tags do not require to be equipped with batteries since they are powered
by the reader if they are within reading distance. Their reading and writing distance,
which depends on the employed communication frequency, can be assumed to be within
a range of meters.
RFID-SLAM is capable of jointly improving multiple trajectories from humans and
robots and requires the autonomous deployment of RFID tags in the environment for
building a network of connected locations. Thereby the pose of each human is automatically
tracked by PDR (Pedestrian Dead Reckoning), which recognizes human footsteps
analytically from acceleration patterns. For the purpose of robot tracking, a slippagesensitive
odometry that reduces odometry errors on slippery ground, as for example, if
the robot navigates on rough terrain, has been developed.

6 Chapter 1. Introduction
With the centralized version of RFID-SLAM, humans and robots are estimating the
distances between RFID tags by pose tracking and communicate them to a central station.
The central station successively builds a joint graph from these estimates and corrects
the joint network of all trajectories by minimizing the Mahalanobis distance [Lu
and Milios, 1997], while utilizing RFID transponders for data association [Kleiner et al.,
2006c]. One advantage of RFID-SLAM is that single agents are not required to perform
loops on their routes since those are automatically generated by merging individual
tracks of the team members. Furthermore, DRFID-SLAM, a decentralized version
of RFID-SLAM, is contributed. DRFID-SLAM utilizes information sharing between
humans and robots via the memory of RFIDs. Hence, does not depend on radio communication
directly.
Team coordination can be generally decomposed into task assignment and multiagent
path planning. Whereas task assignment and multi-agent path planning were both
intensively studied as separate problems in the past, there has been only little attention
on solving both of them at once, particularly if large robot teams are involved. This is
mainly due to the fact that the joint state space of the planning problem grows enormously
with the number of robots. However, particularly in destroyed environments
where robots have to overcome narrow passages and obstacles, path coordination is
essential in order to avoid collisions and deadlocks. In this thesis a novel deliberative
approach that reduces significantly the size of the search space by utilizing RFID tags
as coordination points, while solving the problem of task assignment and path planning
simultaneously, is contributed. The method utilizes the same RFID infrastructure as
RFID-SLAM. Hence, global path planning is carried out on a graph topology, which is
computationally cheaper than planning on global grid maps, as it is usually the case.
Furthermore, a local search method that uses the memory of RFIDs for labeling visited
places in the field is contributed. Since data exchange is carried out via the memory
of RFIDs, the method can also be applied if radio communication fails completely. Additionally,
the local approach has the advantage that computational costs do not grow
with the number of robots participating in the search.
To utilize RFID technology in emergency response scenarios has two further advantages:
first, RFID tags that have been distributed into the environment can be used
in a straightforward manner by humans as landmarks to follow routes towards victim
locations and exits, i.e. they do not need to localize themselves within a metric map.
Second, RFID tags can be used by human task forces to store additional information regarding
nearby places for subsequent rescue teams. The idea of labeling locations with
information that is important to the rescue task has already been applied in practice.
During the disaster relief in New Orleans in 2005, rescue task forces marked building
entrances with information concerning, for example, hazardous areas or victims inside
the buildings [FEMA, 2003]. The RFID-based marking of locations is a straight forward
extension of this concept.

1.2. Contribution 7
In order to navigate autonomously in disaster areas, robots need to have an adequate
representation of the environment. For this purpose, trajectories, as for example generated
by pose tracking and RFID-SLAM, can be taken as a basis for subsequently integrating
range measurements from a Laser Range Finder (LRF) into a consistent map.
In the context of emergency response, robots have to cope with rough terrain, e.g. an
unstructured environment containing obstacles of arbitrary shape in three dimensions.
However, standard approaches for mapping, such as occupancy grid maps [Moravec,
1988], lead to a two-dimensional representation of the environment, hence not supplying
an adequate representation in this context. Furthermore, pose tracking cannot
reliably be performed from wheel odometry or scan matching only.
Therefore, an efficient method for building elevation maps in real-time, i.e. to map
the environment while the robot is in motion, is contributed. The proposed approach
tracks the three-dimensional pose of the robot by integrating measurements from an
orientation sensor and a novel method of visual odometry combined with scan matching.
The visual odometry tracks salient features with the KLT feature tracker [Tomasi
and Kanade, 1991] over images taken by the camera, and computes from the tracked
features the translation of the robot. Furthermore, the three-dimensional pose is updated
from height observations that have been registered on the map. Given the threedimensional
pose, the height value of each map cell is estimated by integrating readings
from a downwards tilted LRF. Due to the integration of the full three-dimensional pose,
the method allows to create elevation maps online while the robot traverses rough terrain,
as for example while driving over ramps and stairs.
To search for victims in disaster areas requires robots to perform subtasks, such as
victim detection, partially or even fully autonomous. Due to the real-time constraint in
rescue-like applications, only fast computable techniques are admissible. However, the
detection rate of fast classifiers, such as color thresholding, motion detection, and shape
detection turns out to be moderate, since they are typically tuned for specific features,
as for example a specific face color at a particular illumination. Hence, in environments
containing many diverse objects, they tend to produce a large number of detections,
from which in the worst case most are false-positives, i.e. objects that are wrongly
recognized as victims. One solution to this problem is to combine local evidences, i.e.
evidences that are close to each other in the real world, and to reason on their true class
label with respect to their neighborhood relations. Markov Random Fields (MRFs)
provides a probabilistic framework for representing such local dependencies. However,
inference in MRFs is computational expensive, and hence not generally applicable in
real-time.
A novel approach for the genetic optimization of MRF models is contributed, which
determines offline relevant neighborhood relations, for example the relevance of the
relation between heat and motion, with respect to the data. These preselected types are
then utilized for generating MRF models during runtime. First, the vertices of the MRF
graph are constructed from the output of the weak classifiers. Second, edges between

8 Chapter 1. Introduction
these nodes are added if the specific type of nodes can be connected by an edge type
that was selected during the optimization procedure. One advantage of the proposed
approach is that it can easily be extended for other sources of evidence, such as CO 2
and audio noise detection.
In order to develop an overall strategy for disaster mitigation, information from single
units, such as detected victims and traveled trajectories, has to be collected and analyzed
centrally, e.g. at a central command post. Only a consistent view on the disaster situation
allows to efficiently coordinate and schedule rescue missions performed by human and
robot teams in the field. Therefore, solutions are needed that increase the situation
awareness of an incident commander from a central perspective, and support the process
of decision making.
The advantage of central data fusion is that the generated model of the disaster area
can be taken as input for algorithms that analyze the current situation and compute
optimal strategies for disaster mitigation and first responder risk reduction. Research
in the field of Multi-Agent Systems (MAS) offers a rich set of solutions for this purpose.
Since in a MAS the number of situations and joint actions performed by agents
is typically large, close-to-disaster-reality simulations are needed for evaluating these
methods. The RoboCup Rescue simulation system aims at simulating large-scale disasters
and exploring new ways for the autonomous coordination of rescue teams [Kitano
et al., 1999]. Here the goal is to provide a platform for developing software agents that
react to disaster situations by coordinating police, ambulance and firefighters. This goal
leads to challenges like the coordination of heterogeneous teams of hundreds of agents,
the exploration of large-scale environments in order to localize victims, as well as the
scheduling of time-critical rescue missions.
Systems that integrated human resources via agent proxies into a MAS are referred
to as human-in-the-loop architectures. A human-in-the-loop architecture is contributed
that integrates first responders via wearable computing into the RoboCup Rescue simulation
system. The proposed wearable device allows to acquire disaster relevant data.
Thereby locations of first responders can either be tracked by GPS positioning or PDR
combined with RFID-SLAM. Finally, the acquisition of real-world data and the coordination
of disaster mitigation in large-scale environments from a central-command-postperspective,
is demonstrated.
Methods introduced in this thesis were extensively evaluated in outdoor environments,
as well as in official Urban Search and Rescue (USAR) test arenas designed
by the National Institute of Standards and Technology (NIST) [Jacoff et al., 2001]. Furthermore,
the proposed methods were an integral part of robotic or multi-agent systems
that won in total more than 10 times the first prize at international competitions, such as
the RoboCup world championships. Presented results show that these methods perform
robustly and efficiently in the utilized benchmark scenarios.

1.3. RoboCup Rescue 9
1.3 RoboCup Rescue
Disaster mitigation is a serious social issue which involves a very large number of heterogeneous
units in hostile environment. Due to the close-to-reality requirement for approaches
intended to be applied in this domain, it is necessary to introduce benchmarks
for evaluating them before they are deployed. After the Great Hanshin Earthquake 1996
in Kobe, the Japanese government decided to promote research related to problems during
large-scale urban disasters. One of the outcomes of this initiative was the RoboCup
Rescue competition. Using the same successful method of competition-based benchmarks
that the Robot Soccer competition was applying, the RoboCup Federation added
two new competitions to RoboCup in 2001, the Rescue Robot League and the Rescue
Simulation League, with the main goals of promoting research and development of
robot and multi-agent technology for disaster mitigation, and their thorough evaluation
regarding real-world deployments [Tadokoro et al., 2000].
In the Rescue Robot league, physical robots are designed and tested in simulated
disaster situations which are mainly designed by the U.S. National Institute for Standards
and Technology (NIST) [Jacoff et al., 2003, Messina et al., 2005]. There, the aim
of robots is to map an unknown environment and provide information about simulated
disaster victims, such as their location and situation, their simulated medical condition
and other helpful indications such as ID tags. The Rescue simulation league aims at
simulating large-scale disasters and exploring new ways for the autonomous coordination
of rescue teams [Kitano et al., 1999]. The goal of teams participating in the
competition is to provide a software system that reacts to a simulated disaster situation
by coordinating a group of agents. This goal leads to challenges like the coordination of
heterogeneous teams, the exploration of a large-scale environment in order to localize
victims, as well as the scheduling of time-critical rescue missions. Agents have only
a limited amount of communication bandwidth they can use to coordinate with each
other. The problem cannot be addressed by a single entity, but has to be solved by a
true multi-agent system. Moreover, the simulated environment is highly dynamic and
only partially observable by a single agent. Agents have to plan and decide their actions
asynchronously in real-time.
Both leagues change elements of the competition and scoring functions from competition
to competition to foster the development of new capabilities and sustain a continuous
progress towards obtaining real-world usable systems, which is the long-term
goal of the leagues. Comparing the results of RoboCup Rescue over the last years, one
notices a remarkable progress in specific areas. Whereas early platforms in the robot
league were mainly wheel-based, later developments included innovative tracked-based
robots capable of climbing stairs and ramps by controlling three independent track segments
[Koyanagi et al., 2006]. Approaches for Simultaneous Localization And Mapping
(SLAM) have been improved from simple compass-based map integration [Birk,
2003] towards mapping of unstructured environments in three dimensions [Surmann

10 Chapter 1. Introduction
et al., 2004]. Furthermore, purely human operator based systems have been partially
replaced by autonomous robot architectures [Kleiner et al., 2006b].
The simulation league developed techniques for multi-agent strategy planning and
team coordination in an inherently distributed rescue effort [Kleiner and Ziparo, 2006,
Kleiner et al., 2004], as well as on the development of information infrastructures and
decision support systems for enabling incident commanders to efficiently coordinate
rescue teams in the field. For example, Schurr introduced a system based on software
developed in the Rescue Simulation league for the training and support of incident
commanders in Los Angeles [Schurr et al., 2005].
1.4 Structure of the thesis
The thesis is structured as follows. In Chapter 2 design criteria for rescue robots are discussed
and hardware platforms that were developed for the evaluation of the proposed
methods, are introduced. In Chapter 3 methods for pose tracking of humans and robots
in USAR-like scenarios are introduced and evaluated. The RFID technology-based
SLAM approach (RFID-SLAM), which is based on pose tracking methods described
in the previous chapter, is introduced and evaluated in Chapter 4. Here the problems of
SLAM by robots, SLAM by humans, and the jointly mapping by humans and robots
are addressed. Furthermore, a decentralized version of RFID-SLAM is presented.
An efficient method for building elevation maps in real-time is presented and evaluated
in Chapter 5. This method is particularly tailored for online processing, thus
facilitating autonomous robot behaviors while navigating on rough terrain. Chapter 6
introduces and evaluates a novel method for coordinating large robot teams in USAR
scenarios. The method utilizes the memory of RFIDs for enabling communication-free
team coordination. A method for detecting humans in USAR environments is introduced
and evaluated in Chapter 7. The introduced approach allows real-time detection
of victims based on optimized Markov Random Fields (MRFs). Finally, in Chapter 8
we discuss an interface for human responders allowing the integration of data that has
been collected in the field into a Multi-Agent System (MAS). Furthermore, based on
the central view from an incident commander generated from this data, methods for
coordinating emergency responders are introduced.
1.5 Relevant publications and awards
Parts of the thesis have been published in the following journal articles, conference,
symposium, and workshop proceedings:

1.5. Relevant publications and awards 11
Journals and Book Chapters
• Real-time Localization and Elevation Mapping within Urban Search and
Rescue Scenarios joint work with C. Dornhege [Kleiner and Dornhege, 2007]
(contained in Chapter 3, Chapter 4, and Chapter 5).
• Towards heterogeneous robot teams for disaster mitigation: Results and
Performance Metrics from RoboCup Rescue joint work with S.Balakirsky,
S.Carpin, M. Lewis, A. Visser, J.Wang, and V. A. Ziparo [Balakirsky et al., 2007]
(contained in Chapter 6).
• Game AI: The shrinking gap between computer games and AI systems [Kleiner,
2005].
Conference Papers
• Genetic MRF model optimization for real-time victim detection in Search
and Rescue joint work with R. Kümmerle [Kleiner and Kümmerle, 2007] (contained
in Chapter 7).
• Decentralized SLAM for Pedestrians without direct Communication joint
work with D. Sun [Kleiner and Sun, 2007] (contained in Chapter 3 and Chapter
4).
• Behavior maps for online planning of obstacle negotiation and climbing on
rough terrain joint work with C. Dornhege [Dornhege and Kleiner, 2007a] (partially
contained in Chapter 5).
• Cooperative Exploration for USAR Robots with Indirect Communication
joint work with V.A. Ziparo, L. Marchetti, A. Farinelli, and D. Nardi [Ziparo
et al., 2007a] (contained in Chapter 6).
• RFID-Based Exploration for Large Robot Teams joint work with V.A. Ziparo,
B. Nebel, and D. Nardi [Ziparo et al., 2007b] (contained in Chapter 6).
• RFID Technology-based Exploration and SLAM for Search And Rescue joint
work with J. Prediger, and B. Nebel [Kleiner et al., 2006c] (contained in Chapter
4).
• Successful Search and Rescue in Simulated Disaster Areas joint work with
M. Brenner, T. Bräuer, C. Dornhege, M. Göbelbecker, M. Luber, J. Prediger, J.
Stückler, and B. Nebel [Kleiner et al., 2005a] (contained in Chapter 8).
• Approaching Urban Disaster Reality: The ResQ Firesimulator joint work
with T. A. Nüssle, and M. Brenner [Nüssle et al., 2004].

12 Chapter 1. Introduction
Workshop Papers
• Mapping disaster areas jointly: RFID-Coordinated SLAM by Humans and
Robots joint work with C. Dornhege and D. Sun [Kleiner et al., 2007] (contained
in Chapter 4).
• Towards the Integration of Real-Time Real-World Data in Urban Search and
Rescue Simulation joint work with H. Kenn [Kenn and Kleiner, 2007] (contained
in Chapter 8).
• Wearable computing meets multiagent systems: A real-world interface for
the RoboCupRescue simulation platform joint work with N. Behrens, and H.
Kenn [Kleiner et al., 2006a] (contained in Chapter 8).
• Visual Odometry for Tracked Vehicles joint work with C. Dornhege [Dornhege
and Kleiner, 2006] (contained in Chapter 3).
Team Description Papers
• RoboCupRescue - Simulation League Team RescueRobots Freiburg (Germany)
joint work with V.A. Ziparo [Kleiner and Ziparo, 2006] (contained in
Chapter 6).
• RoboCupRescue - Robot League Team RescueRobots Freiburg (Germany)
joint work with C. Dornhege, R. Kümmerle, M. Ruhnke, B. Steder, B. Nebel, P.
Doherty, M. Wzorek, P. Rudol, G. Conte, S. Durante, and D. Lundstrom [Kleiner
et al., 2006b] (contained in Chapter 2).
• RoboCupRescue - Robot League Team RescueRobots Freiburg (Germany)
joint work with B. Steder, C. Dornhege, D. Höfler, D. Meyer-Delius, J. Prediger,
J. Stückler, K. Glogowski, M. Thurner, M. Luber, M. Schnell, R. Kümmerle, T.
Burk, T. Bräuer, and B. Nebel [Kleiner et al., 2005b] (contained in Chapter 2).
• ResQ Freiburg: Team Description and Evaluation joint work with M. Brenner,
T. Bräuer, C. Dornhege, M. Göbelbecker, M. Luber, J. Prediger, and J. Stückler
[Kleiner et al., 2004] (contained in Chapter 8).
Awards
Together with teams of students and colleagues we were able to gain the following
awards at international competitions:
• In the RoboCup Mid-sized league (with CS Freiburg):

1.5. Relevant publications and awards 13
– 1st place at the GermanOpen 2001
– 1st place at the RoboCup world championship 2001 in Seattle
– 2nd place at the GermanOpen 2002
• In the RoboCupRescue Simulation league (with ResQ Freiburg):
– 1st place at the GermanOpen 2003
– 1st place at the GermanOpen 2004
– 1st place at the RoboCup world championship 2004 in Lisboa
– Winner of the Infrastructure Competition at the RoboCup world championship
2004 in Lisboa
• In the RoboCupRescue Robot league (with RescueRobots Freiburg):
– 2nd place at the GermanOpen 2005
– Winner of the mobility award at the GermanOpen 2005 (Together with AIS
Fraunhoffer)
– 1st place of the "Best in class Autonomy" competition at the RoboCup world
championship 2005 in Osaka
– 1st place of the "Best in class Autonomy" competition at the RoboCup world
championship 2006 in Bremen
• In the RoboCupRescue Simulation league (Virtual Robots) (with Vittorio Ziparo)
– 1st place at the RoboCup world championship 2006 in Bremen
• In the RoboCupRescue Simulation league (Infrastructure competition) (with
Nils Behrens and Holger Kenn)
– Winner of the Infrastructure Competition at the RoboCup world championship
2006 in Bremen
• At the Sick Robot Day 2007, held by the SICK GmbH (with Jörg Müller and
Christian Dornhege)
– 1st place of the indoor competition

2 Rescue Robotics Hardware
In this chapter, hardware components for rescue robotics, which have been developed
or used for experimental evaluations in this thesis, are introduced. On the one hand,
design criteria for robot platforms operating in rescue scenarios are discussed. On the
other hand, sensor devices, such as Laser Range Finders (LRFs) and IR cameras are
introduced. Furthermore, RFID technology, which serves as a basis for algorithms
proposed in this thesis, is discussed.
Finally, two robot platforms that have been designed and developed by Rescue Robots
Freiburg at the University of Freiburg, and methods for Human Robot Interaction (HRI),
are described. HRI methods introduced in this chapter are designed for facilitating the
control of a team of heterogeneous robots by a single operator. The introduced robot
platforms Zerg and Lurker are tailored for different types of tasks. Whereas the Zerg
robot aims at large-scale exploration within a robot team, the Lurker robot has been
particularly designed for overcoming rough terrain, such as climbing stairs and ramps.
The remainder of this chapter is structured as follows. In Section 2.1 general design
criteria for rescue robot platforms are discussed. Sensors for navigation and perception
that have been utilized for experiments proposed in this thesis, are introduced in
Section 2.2 and Section 2.3, respectively. Section 2.4 introduces RFID technology and
discusses its importance to rescue robotics. Finally, in Section 2.5 the developed robot
platforms are proposed.
2.1 Design criteria for robot platforms
Designing robot platforms for Urban Search and Rescue (USAR) is a challenging problem.
A detailed analysis of the requirements for the performance of robots that can
support USAR roles and tasks has been undertaken by the US authorities NIST (National
Institute of Standards and Technology) and DHS (Department of Homeland Security)
[Messina et al., 2005]. They acquired expert knowledge by polling members
from twenty FEMA (Federal Emergency Management Agency) task forces and the US
National Guard. Among other roles for USAR robots, they identified reconnaissance
and primary search, as the two highest priorities for applying robots. The final result of
the survey has been utilized to establish performance metrics for USAR robots.
Among other things, the platform has to fulfill the following criteria. First, any tasks
in disaster areas, such as finding victims, demands a high degree of versatility from the
15

16 Chapter 2. Rescue Robotics Hardware
(a) (b) (c)
Figure 2.1: Rescue Robots during the NIST “Response Robot Exercise” evaluated by
performance metrics for their USAR usability: (a) IRobot PackBot EOD with (b) teleoperation
console. (c) The AirRobot tele-operated for flying reconnaissance missions.
Courtesy of Raymond Sheh.
robots. For example, there are places only reachable by climbing, spots only accessible
through small openings, and regions only observable from the air. Robots operating
in such environments have to be designed with an adequate degree of mobility and
size. Ideally, they compose a heterogeneous team combining very different individual
capabilities, which is therefore capable of overcoming very different types of terrain.
Second, robots and humans must work together in order to successfully accomplish
their mission. Hence, robot platforms have to be designed that they are fitting into
existing rescue operations, i.e. they have to actively support the work of first responders
operating in the field. For example, a platform has to be real-time controllable within
confined spaces no less than 30 meters far from the operator and being suitable for longterm
deployment of at least 12 hours with rechargeable battery packs. Furthermore, the
human-robot interface has to be intuitive usable by untrained personal via a limitedfunction
controller and a user-friendly GUI (Graphical User Interface). Third, robots
have to be designed for rugged use and being affordable by rescue organizations.
Figures 2.1 and 2.2 show robots evaluated during the NIST response robot exercise in
Gaithersburg (2006), where a wide range of diverse robot platforms participated. Figure
2.1 (a) depicts the PackBot EOD, a robot equipped with a manipulator allowing to
collect information, such as video images, thermo image signatures, and CO 2 emission,
at locations that are difficult to access. Figure 2.1 (c) depicts the AirBot. This platform
enables first responders to get an overview of a large terrain, or to seek access into

2.1. Design criteria for robot platforms 17
(a)
(b)
Figure 2.2: Rescue Robots during the NIST “Response Robot Exercise” evaluated by
performance metrics for their USAR usability: (a) The Soryu snake robot, and (b) the
BOZ robot lifting a smaller DragonRunner robot to an upper location. Courtesy of
Raymond Sheh.
(a)
(b)
Figure 2.3: Legged rescue robots: (a) The R-Hex robot. (b) Walking machines from
the University of Bremen. Courtesy of Raymond Sheh (a) and Adam Jacoff (b).
buildings from the roof. The Soryu snake (Figure 2.2 (a)) has been developed particularly
for accessing collapsed buildings through interstices in floors. Furthermore, as
depicted in Figure 2.2 (b) there might be tasks that are optimally solved by teamwork
between robots. Finally, Figure 2.3 depicts legged rescue robots that are designed for
overcoming rough terrain. Their design is mainly inspired by insects, such as ants and
cockroaches, which show great mobility capabilities in their natural environment.

18 Chapter 2. Rescue Robotics Hardware
Up to now, robot autonomy, i.e. robots controlled by Artificial Intelligence methods,
is only rarely found on truly fieldable systems. Rescue robots have been exclusively deployed
under tedious human supervision. For example, single control commands, such
as positioning a manipulator or moving a robot to a target location, are issued from a
tele-operation console, as shown in Figure 2.1 (b). This causes a high burden on human
operators and limits the number of robots they can deploy at the same time. Therefore,
the ultimate goal is to increase the level of autonomy of rescue robots step-wise, thereby
enabling a single operator to control a large team by a small set of high level commands,
such as assigning an area that has to be explored, whereas the robots autonomously navigate
to their target locations. This goal requires that robots are equipped with sensors
for localization and mapping, such as Laser Range Finders (LRFs), and Inertial Measurement
Units (IMUs), as well as sufficient computational power.
Requirements for autonomy have a great impact on the design of robot platforms. Unfortunately,
most commercially available robot platforms are not particularly designed
for autonomy support. On the one hand, an increase of computational capacities causes
an increase in the robot’s weight, which is, for example, a crucial problem in case of
the AirBot robot platform. On the other hand, sensors for localization require a special
mounting, e.g. LRFs require a free line of sight within an angle of 180 ◦ or even 270 ◦ ,
and IMUs have to be mounted outside magnetic fields, as they are for example caused
by the motors. In turn, the implementation of these constraints affects the mobility of
the platform significantly, or even demands its complete re-design. In Section 2.5 two
robot platforms that were particularly developed for autonomy missions are described.
2.2 Sensors for navigation
2.2.1 Inertial Measurement Unit (IMU)
An Inertial Measurement Unit (IMU) is a closed system that is used to detect orientation
and motion of a vehicle or human. It typically contains a combination of three
accelerometers and three angular rate sensors (gyroscopes), which are placed such that
their measuring axes are orthogonal to each other. Accelerometers measure the acceleration
by determining inertia forces acting on a body, whereas a gyroscope measures
changes of the body’s orientation based on the principle of conservation of angular momentum.
Gyroscope measurements are unreliable in the long term since they are subject
to drift, which can partially be reduced if the IMU has a temperature sensor.
The combination of an IMU with a magnetometer (compass) or a Global Navigation
Satellite System (GNSS) is referred to as Attitude and Heading Reference System
(AHRS). Such systems typically contain a Kalman filter for fusing the data from multiple
sensor sources, yielding an orientation vector that is comparably stable towards
sensor drift and magnetic perturbations.

2.2. Sensors for navigation 19
IMUs have been utilized traditionally for aircraft navigation. Due to the increasing
miniaturization and decreasing price of these devices, they are more and more applied
to car and robot navigation. More recent applications include human motion capture in
the context of sports technology, computer animation, and support for the navigation of
blind people.
Xsens MTi/MTx
The MTi is a miniature, gyro-enhanced AHRS and combines a tri-axial accelerometer
and a tri-axial gyroscope with a tri-axial magnetometer (see Figure 2.4). From these
internal sensors, the unit computes a three-dimensional orientation vector with a signal
processor, which is updated internally at 512 Hz, and defined by the three Euler
angles yaw (azimuth or compass angle), roll (bank), and pitch (elevation). Due to the
simultaneous integration of gyro and compass data, the orientation vector is drift-free
and stable towards minor perturbations caused by external magnetic sources. The ori-
Figure 2.4: Two Inertial Measurement Units (IMUs): The MTi and MTx from Xsens.
Courtesy of Xsens.
entation vector, the calibrated 3D acceleration, the 3D rate of turn (rate gyro), and the
3D earth-magnetic field data, are accessible via a RS-232 interface at 120 Hz. Under
static conditions, i.e. when the senor is not in motion, the roll and pitch components
of the vector have an accuracy of

20 Chapter 2. Rescue Robotics Hardware
AMD (Adapt to Magnetic Disturbances), which can be activated in environments containing
heavy magnetic fields.
The MTx from Xsens is a slightly modified version of the MTi, which is designed
particularly for capturing human motion. The device mainly differs from the MTx by
the casing’s shape and weight and that accelerations are captured at a higher frequency.
2.2.2 Laser Range Finder (LRF)
Laser Range Finders (LRFs) are commonly used in robotics for localization and mapping
due to their high accuracy and fast data data rate. They measure the distances to
surrounding objects according to the Time Of Flight (TOF) principle. The distance to
an object is computed from the time it takes to detect the laser beam that has been emitted
by the sensor and reflected by the object. Typically, beams at half-degree resolution
are emitted in a Field Of View (FOV) of 180 ◦ . A widely used model is the LMS200
from Sick since this sensor is extremely robust and capable to detect objects within a
range of 80 meters. However, this device greatly influences the mobility of a robot platform
due to its heavy weight of 4.5 kg. A lightweight solution is the URG-04LX from
Hokuyo, which weighs only 160 g. However, its limited range of only 4 meters makes
it difficult to apply algorithms for localization and mapping since these require a rich
set of features. A maximal range of 4 meters leads in many cases, as for example in
outdoor situations, to a large number of far readings, i.e. measurements without reflections,
whereas in narrow indoor situations, as for example the interior of a rubble pile,
are appropriate scenarios for its deployment. The S300 from Sick offers a compromise
between both sensors since it is much lighter than the LMS200, but only measures distances
up to 30 meters. Table 2.1 summarizes the technical specifications of LRFs that
have been utilized during experiments introduced in this thesis.
Sick LMS200 Sick S300 Hokuyo URG-04LX
Weight 4500 g 1200 g 160 g
Volume ≈ 20 cm 3 ≈ 15 cm 3 ≈ 5 cm 3
FOV 180 ◦ 270 ◦ 240 ◦
Max. Range 80 m 30 m 4 m
Max. Ang. Res. 0.25 ◦ 0.5 ◦ 0.36 ◦
Accuracy ± 15 mm ± 30 mm ± 10 mm
Scans per second 30 20 10
Interface RS-232/RS-422 RS-232/RS-422 RS-232/USB
Table 2.1: Comparison of laser range finders.

2.2. Sensors for navigation 21
2.2.3 Global Navigation Satellite System (GNSS)
GNSS (Global Navigation Satellite System) refers to the general class of satellite-based
positioning systems, such as GPS, GLONASS, and Galileo, whereas the term GPS virtually
refers to the US-American Navstar System only. Since the Global Positioning
System (GPS) is the most commonly used GNSS, it will be described in more detail.
GPS is composed of a nominal constellation of 24 satellites ∗ , where the orbits of
the satellites are arranged in a way that at least six satellites are always within line of
sight from almost anywhere on Earth. GPS positioning is carried out by measuring
the distance between the receiver and four or more GPS satellites. Since the transmission
speed of radio signals is constant (nearly at the speed of light), the distance to a
satellite can be computed from the measured time delay between the transmission and
reception of the radio signal, i.e. from the signal’s propagation time. This procedure requires
extremely accurate clocks on both the receivers and transmitters side. However,
receivers are typically equipped with low-cost crystal oscillator-based clocks since accurate
atomic clocks, as they are used by the satellites, are too expensive for consumer
devices. Therefore, the clock of the receiver has to be synchronized continuously with
a satellite clock, requiring to track one satellite only for this purpose.
Each satellite transmits an individual Pseudo Random Code (PRC) from which the
identity of the satellite and the time delay of the signal can be determined. Additional
information, such as the exact locations of the satellites, is modulated on this signal
and decoded by the receiver. From the known locations and distances of satellites the
location of the receiver can be deduced by a technique known as trilateration. The
technique can be visualized by the intersection of imaginary spheres, which have centers
at each satellite location, and radii equal to the corresponding distance measurements.
Theoretically, four satellite locations and distances are necessary to deduce the threedimensional
location of the receiver. However, under the assumption that the GPS
receiver is located on the earth’s surface, the correct location can be disambiguated
reliably from only three satellites since one of the two concluded locations is generally
outside the satellite orbit.
The farther the tracked satellites are from each other, the better the accuracy of the
estimated position since the spheres are intersected by more acute angles. A value
expressing the quality of the current satellite constellation is the Horizontal Dilution
of Precision HDOP value, which is provided by most GPS receivers. In summary,
the location of the receiver, denoted by the three unknown parameters longitude (East
coordinate) and latitude (North coordinate), and altitude, can be deduced successfully
from four satellites, including one satellite for synchronization of the receiver clock.
If there are sufficient satellites within line of sight of the receiver, atmospheric effects
have the biggest influence on the accuracy of GPS positioning. Atmospheric effects
∗ Note that since January 2007 there are 29 broadcasting satellite in the GPS constellation in order to
increase the positioning accuracy by redundancy.

22 Chapter 2. Rescue Robotics Hardware
might cause a variation of the radio signal’s velocity. Particularly the ionosphere, which
is 50 - 500 km located above the surface of the earth, contains ionized particles that influence
propagation speeds. The fact that these effects are comparably durable within a
particular region makes it possible to compensate them on the receiver’s side by utilizing
a reference measurement that has been taken in the same region. This procedure is
known as Differential GPS (DGPS), and is carried out by stations, which periodically
determine the error between a local range measurement to a satellite, and the range that
has been computed from the accurately known positions of satellite and station.
The computed correction can be used for improving position accuracy of all receivers
located in the station’s region. The correction can be transmitted directly to the receiver
by sending it over radio (which requires that the receiver is able to receive on an additional
channel), or it can be received alternatively via an Internet connection.
Recently there has been a development of Satellite Based Augmentation Systems
(SBAS) that transmit the corrections on the same frequency as the positioning satellites,
having the effect that no additional receiver hardware is required. Correction data
and positioning data are distinguished from each other by the Code Division Multiplex
Access (CDMA) procedure. The European Geostationary Navigation Overlay Service
(EGNOS), which mainly covers areas in Europe, consists of three geostationary satellites
and a network of ground stations. EGNOS increases positioning accuracy of GPS,
GLONASS, and Galileo from 10 - 20 meters down to 1 - 3 meters. Comparable systems,
which are compatible with EGNOS, are in the United States the Wide Area Augmentation
System (WAAS), and in Japan the MTSAT Space-based Augmentation System
(MSAT).
Experiments presented in this work have been carried out with the GPSlim236 GPS
receiver from Holux, which is equipped with SIRFstar III technology. The receiver
allows to track up to 20 satellites at an update rate of 1 Hz and has a position accuracy
of 5 - 25 meters without DGPS. Furthermore, the receiver is able to process data from
the EGNOS system, improving the horizontal position accuracy to

2.2. Sensors for navigation 23
Figure 2.5: Map of Europe showing the latitude and longitude zones of the Universal
Transverse Mercator (UTM) projection from 29 S to 38 W. Courtesy of the Demis
company.
of 6 ◦ longitude (800 km), where each zone is centered over a meridian of longitude.
Starting from the date line, zones are numbered from 1 to 60 increasing in an easterly
direction. Furthermore, each vertical zone is separated into 20 latitude zones of a height
of 8 ◦ , which are lettered starting from "C" at 80 ◦ S, increasing up until "X". Figure 2.5
shows the corresponding zones for Europe.
Within each Zone, locations are addressed by an easting and northing coordinate pair
(x, y), with the point of origin at the intersection of the equator with the central meridian
of the zone. In order to avoid negative easting numbers, an offset of 500,000 meters is
added to each easting. Hence, locations east of the zone’s central meridian have an
easting value above 500,000 meters, whereas locations western of it have a value below
500,000 meters.
A similar procedure is applied to northing numbers. Northings in the northern hemisphere
increase as going northward from the equator, which has an initial northing value
of 0 meters, whereas in the southern hemisphere, northings decrease as going southward
from the equator, which has then an initial value of 10,000,000 meters in order
to avoid negative numbers. Therefore, it theoretically suffices to describe a zone by its
vertical zone number and an indication whether the zone is located in the northern or
in the southern hemisphere. For example, my office at the Computer Science Department
at the University of Freiburg is located at the geographic position 48 ◦ 0’49.03" N,
7 ◦ 50’1.96" E (in decimal 48.013621, 7.83388), which corresponds to the UTM zone
32U with an easting value of 413036 meters and a northing value of 5318471 meters.
Besides the described calculation, the projection handles certain areas differently, as for
example the areas around the poles, and also minimizes distortion by a scaling factor.

24 Chapter 2. Rescue Robotics Hardware
The interested reader might find further information on map conversions in the document
published by J. S. Snyder [Snyder, 1987].
2.3 Sensors for detecting humans
2.3.1 Video and infra-red cameras
Vision sensors are essential for tele-operation since they are required for navigation
as well as for gathering information on victims, such as identifying their state and the
way they are buried. Moreover, they can be utilized for autonomous victim detection.
This is conducted by applying image processing techniques on data generated from
video and infra-red cameras. However, in the context of emergency response this is
a challenging problem since only little assumptions can be made regarding color and
shape of human body parts. Infra-red cameras, which are capable of detecting body
heat, might misinterpret other heat sources, such as radiators, engines, and fire.
(a) (b) (c)
Figure 2.6: The CCD cameras (a) Sony DFW-500, and (b) Logitech QuickCam 4000
Pro. (c) The infra-red camera Thermal-Eye 3600AS.
Figure 2.6 shows camera models that have been utilized for victim identification in
this thesis. The Sony DFW-V500 camera incorporates a 1/3 type Progressive Scan
Wfine CCD and an IEEE-1394 (Fire Wire) interface. Images are captured with 30
frames/sec at a maximal resolution of 640 × 480 pixels in the YUV 4:2:2 format. The
camera has a weight of 305 g and requires an external power source. It offers highquality
images that are adequate for applying image processing techniques.
The Logitech QuickCam 4000 Pro also captures video images at 640×480 pixels, and
can be connected via USB to a computer. This camera has been used for tele-operation
and visual odometry (see Section 3.4). Due to its lightweight of less than 50 g (without
housing) it can be mounted on a robot easily. Furthermore, the original lens of the
camera has been replaced with a wide-angle lens yielding a FOV of 74 ◦ . However, the
quality of the images turned out to be non-adequate for victim detection.

2.3. Sensors for detecting humans 25
The ThermalEye 3600AS Infra-Red (IR) camera allows to detect human beings by
their temperature within a Field Of View (FOV) of ≈ 50 ◦ × 37 ◦ , and a distance of up
to 100 meters with a specified operating Temperature of −20 ◦ C to 85 ◦ C. The camera
captures images with 30 frames/sec at a maximal resolution of 160 × 120 pixels, which
are provided by an analog PAL signal that has to be digitized with additional hardware.
Due to the modest power consumption of only 1.2 W and a weight of only 67.5 g, the
camera can be mounted on nearly every robot platform. Each pixel of the camera image
contains a gray value which is proportional to the measured heat at the real-world
location corresponding to the pixel.
Before camera images can be used for victim detection, the camera has to be calibrated
with respect to its intrinsic parameters, such as the focal length and the radial
lens distortion. On color cameras these parameters are usually found by applying an
analytical method [Tsai, 1986] that utilizes pixel to real-world correspondences as input.
These correspondences are determined from pictures taken of a test pattern, such
as the printout of a chess board [Bradski, 2000], with known feature distances. Then,
relative displacements of features detected in the images, and those displacements measured
on the test pattern, are brought into relation for computing a conversion function.
In case of IR camera calibration, it is necessary to generate a test pattern that also appears
on thermo images. This can be achieved by taking images from a heat reflecting
metal plate covered with quadratic isolation patches in a chess board-like manner. The
resulting images can then be processed by the calibration procedure in the same way
as if taken from a color camera. Finally, the calibration procedure allows to determine
the real-world distance and angle of features detected in the image. However, accurate
depth information can only be obtained by stereo camera systems, since depth computation
from monocular cameras requires that observed objects are located at the same
hight.
2.3.2 3D Laser Range Finder (LRF)
In Rescue Robotics, the task-relevant environment is three-dimensional, and thus requires
highly developed sensors for autonomous operation. For example, the detection
of structures relevant for navigation, such as ramps, stairs, and stepfields, makes perception
of three-dimensional objects indispensable. Also for the task of victim detection
image processing techniques are not always sufficient on their own. Accurate depth
measurements are needed in order to correctly group vision features according to their
spacial neighborhood, and to reliably distinguish between background and foreground.
Therefore, we developed a light-weight 3D Laser Range Finder (LRF) device for structure
detection and mapping in the rescue domain. A picture of the sensor, and a 3D
scan taken by this device can be seen in Figure 2.7. The LRF sensor is rotated by a
strong servo that allows fast and accurate vertical positioning of the device. The device
can be rotated vertically by more than 90 degrees, which is sufficient to generate 3D

26 Chapter 2. Rescue Robotics Hardware
(a)
(b)
Figure 2.7: A hand-crafted light-weight 3D Laser Range Finder (LRF) device. (a) A
model of the 3D LRF. (b) 3D scan taken in front of stepfield and baby doll.
scans from humans, and objects, such as stairs, ramps, and stepfields. The design differs
from other 3D scanners in that it can be implemented by a simple “U-shaped” piece
of metal and a servo, which altogether can be produced at approximately 100 USD (see
Section 10.2.1).
2.3.3 Audio and CO 2 sensors
The locations of survivors after a disaster can be determined by audio through utilizing
two microphones mounted with a fixed distance between them on the robot platform
(see Figure 2.8 (a)). Given an audio source left, right or between both microphones,
it is possible to measure the time difference, i.e. phase shift, between both signals.
Kenn et al. use differential time of flight measurements through energy cross-spectrum
evaluation of the sound signals for detecting the angular direction to multiple sound
sources [Kenn and Pfeil, 2006]. They integrate all measurements into an occupancy grid
map for localizing the victims in a global map. We utilized the Crosspower Spectrum
Phase (CSP) approach, which allows to calculate the phase shift of both signals based
on the Fourier transformation [Giuliani et al., 1994, Bennewitz et al., 2005, Omologo
and Svaizer, 1996, Svaizer et al., 1997]. As shown in Figure 2.9, the bearing of the
sound source can be successfully determined, even for different kinds of noise. To
locate victims under harsh environmental conditions, e.g. in a noisy environment, is still
challenging. However, the detection of audio signals close to victim locations provides
a good indication of the victim’s state, e.g. it might indicate that the victim is still alive
and thus that a rescue mission is worth the effort.
Gases, such as CO 2 , can be detected by the use of IR measuring after the NDIR (Non
Dispersive Infra-Red) principle. Within a small tube, gases are absorbing infra-red
light at a particular frequency. By shining an infra-red beam through the tube contain-

2.3. Sensors for detecting humans 27
(a)
(b)
Figure 2.8: The Zerg robot (a) equipped with two microphones for victim detection,
and (b) measuring with a Telaire 7001 sensor CO 2 emission of a victim during a TV
recording in Freiburg.
Count
Baby doll Noise
30
25
20
15
10
5
0
-100 -80 -60 -40 -20 0 20 40 60 80 100
Bearing [Deg]
(a)
Count
White Noise
180
160
140
120
100
80
60
40
20
0
-100 -80 -60 -40 -20 0 20 40 60 80 100
Bearing [Deg]
(b)
Figure 2.9: Examples for sound source detection by the Crosspower Spectrum Phase
(CSP) approach. Both sound sources were located at a bearing of +30 degree. (a)
Sporadic noise from a baby doll, and (b) continuous white noise are mainly detected at
the right bearing.
ing CO 2 , and measuring the amount of infra-red absorbed by the gas at the necessary
wavelength, a NDIR detector is able to measure the volumetric concentration of the gas.
Different kinds of wave lengths from IR light can be used for the analysis of different
gases. The gas is either pumped or diffused into a tube, and the electronics measures
the absorption of the characteristic wavelength of light. One drawback of this measurement
principle is its slow response time for detecting a change of the CO 2 concentration.
Therefore, robot navigation has to be interrupted for gaining reliable measurements.
We utilized the Telaire 7001 NDIR carbon dioxide sensor, which measures CO 2 con-

28 Chapter 2. Rescue Robotics Hardware
centrations in the range of 0 to 4000 ppm at an accuracy of 50 ppm, and response time
of < 60 seconds for 90% of the step change (see Figure 2.8 (b)).
The application of CO 2 concentration measurements in the context of urban search
and rescue has been studied [Takashi and Hiroshi, 2003, Nguyen et al., 2004] intensively.
Measurements of the CO 2 concentration can be utilized for survivor search.
However, due to the volatility of CO 2 gas, particularly if the concentration is very low,
survivor localization is possible under special conditions only. Gas sensors are more
promising for determining the victim’s state, and for the detection of hazardous locations,
e.g. by sensing flammable and poisonous gases that can harm first responders
during their missions.
2.4 RFID technology
RFID stands for Radio Frequency Identification, a term that describes small devices
(RFID tags) that use radio signals to exchange data with a reader. The RFID tag transmits
typically a number that is worldwide unique, or at least unique within in the context
of the application. This has the effect that an object, to which the tag is attached to, can
uniquely be identified from a database. In the usual context, RFID tags identify parcels
in a post office or products sold in department stores. For example, since 2005 Wal-
Mart requires from its top 100 suppliers to ship pallets labeled with RFID tags to its
stores. The Metro group started the “Future Store Initiative”, a stepwise application of
RFID technology, ranging from automated stock maintenance in stores, to intelligent
refrigerators in consumer households that monitor the expiration date of food based on
data read from tags attached to the food items. In the long run, their goal is to completely
replace the current bar code system, found on products sold today, by RFID
technology. Furthermore, the size of RFID chips shrinks continuously. For example,
Hitachi introduced 2003 the µ-Chip with a size of 0.4 × 0.4 mm, where the generation
introduced in 2006 reaches a size of 0.15 × 0.15 mm. Their latest version, released in
2007, has a size of 0.05 × 0.05 mm. The progressing miniaturization of RFID technology
has also the effect of reducing manufacturing costs since considerably more chips
can be produced from a single wafer. The small size of RFIDs makes it also possible
to attach them to bank notes [Lange, 2005], or to integrate them into carpets for simplifying
indoor localization of cleaning robots. The German vacuum-cleaner company
Vorwerk introduced in 2006 the “Smart Floor”, a network of RFID chips embedded
into a polyester fabric that can be placed under carpets [Co. KG, 2005]. The recent
advent of the RFID technology leads to a steady increase of smart devices in our environment
that provide digital data. Algorithms and methods applied in this context
are generally recognized under the terms “Sensor networks”, “Ubiquitous Computing”,
and “The Internet of things”.
RFID tags can be classified as active and passive devices depending on whether they

2.4. RFID technology 29
are powered directly from an attached battery, or whether they are powered from an
electromagnetic field emitted by the reading device. This section focuses on passive
devices, which are produced cheaply and in large numbers. They consist mainly of an
antenna, memory, and a sophisticated state machine (or even a microprocessor). The
latter is required if tags handle encrypted communication or implement an anti-collision
protocol. Anti-collision protocols are necessary if the application requires multiple tags
to be read at the same time. RFID tags with anti-collision protocols wait for their turn
when responding to a reader. In contrast to the “1-bit” Electronic Article Surveillance
(EAS) tags, which are commonly found in libraries or stores for theft prevention, these
advanced RFID tags are equipped with a read- and writable memory, which can be used
to store additional information on the object the tag is attached to. It can be assumed that
due to the continuously decreasing price of memory chips, the capacity of RFID tags
will further increase in the future. The company MicroSensys offers passive RFID tags
with up to 512 KBit memory, and develops currently RFIDs with memory capacities of
up to 4 MBit [MicroSensys, 2007].
RFID tags are available for various communication frequencies within the Industrial
Scientific Medical (ISM) band, as for example, at High Frequency (HF) 6.78 MHz,
13.56 MHz and 27.125 MHz, at Ultra High Frequency (UHF) 433.920 MHz, 869 MHz,
915 MHz, and at Microwave 2.45 GHz, 5.8 GHz, and 24.125 GHz. The choice of the
communication frequency is of major importance for the application since the maximal
communication range and communication bandwidth depends on it. The higher the
frequency, the higher the communication bandwidth and communication range. Furthermore,
the maximal range depends on the environment, e.g. the kind of material
located between receiver and tag. For example, metal or water reduces the range significantly.
Short-wave frequencies, as for example 13,56 MHz, lead typically to distances
below three meters and thus are only applicable to objects within short range. Table 2.2
summarizes some RFID frequencies and applications they are utilized for.
Frequency Max. reading range Typical applications
LF (e.g. 100 KHz) 0.5 m Pet identification and
close reads of items
with high water content
HF (e.g. 13.56 MHz) 3 m Building access control,
Product identification
UHF (e.g. 915 MHz) 9 m Boxes and pallets
Microwave (e.g. 2.4 GHz) >10 m Vehicle identification of
all sorts
Table 2.2: Communication frequencies and corresponding maximal reading ranges of
RFID tags and their typical applications (original data from [Bhatt and Glover, 2006]).

30 Chapter 2. Rescue Robotics Hardware
2.4.1 RFIDs in robotic applications
For robotic applications, particularly in the context of localization and mapping, it is
important to compute an estimate on the distance between the RFID tag and the reader.
The Transceiver-Receiver (TR) separation, i.e. the distance between a detected RFID
and the detector, can generally be estimated from the power of the signal. However,
signal propagation in an indoor environment is perturbed by damping and reflections
of radio waves. Since these perturbations depend on the layout of the building, the
construction material used, and the number and type of objects in the building, modeling
the relation between signal path attenuation and TR separation is a challenging problem.
Seidel and Rapport introduced a model for path attenuation prediction that can also be
parameterized for different building types and the number of floors between transceiver
and receiver [Seidel and Rapport, 1992]. This model has been evaluated for frequencies
in the UHF domain, e.g. 914 MHz. RFID implementations operating in this domain are
requiring a line of sight between the tag and the detector. This allows to adopt a simpler
version of the model from Seidel and Rapport, based on the assumption that RFID
detections are not possible through walls [Ziparo et al., 2007a]. The model relates the
signal power P to distance d in the following way:
P(d)[dBm] = p(d 0 )[dBm] − 10n log d d 0
, (2.1)
where P(d 0 ) is the signal power at reference distance d 0 and n denotes the mean path
loss exponent that depends on the structure of the environment. Seidel and Rapport
determined for transmissions at 914 MHz a path loss of 31.7 dB at a reference distance
of 1 meter. Furthermore, they determined for different building types characteristic
values for n and the standard deviation σ of the signal.
TR has also been evaluated in order to improve the security of RFID data transmission
[Hancke and Kuhn, 2005]. RFID tags are vulnerable to relay attacks if they are
used for proximity authentication. Therefore, it is important to read preferably from
tags that are close to the reader, which requires the reader hardware to be aware of the
distance to the sender. Another application is to localize objects with attached RFIDs,
such as books in a library. The basic idea is to build a system that guides the user’s
search by providing feedback on the distance to the searched object [Fishkin and Roy,
2003]. More details on RFID technology are found in [Finkenzeller, 2003].
2.4.2 Importance for Urban Search and Rescue (USAR)
RFID technology offers a great opportunity for disaster mitigation and USAR. The key
problem in this scenario is to get an overview of the disaster, e.g. to be aware of the
locations of first responders while generating a map augmented with disaster-relevant
information, such as victim locations, hazardous areas, and passable openings. How-

2.5. Rescue Robots Freiburg 31
ever, standard localization techniques, such as vision-based structure recognition, and
GPS, are not well suited in these scenarios due to harsh environmental conditions. For
example, smoke and dust particles lead to a limited field of view, and GPS fails completely
within structures made of reinforced concrete. RFID technology can be utilized
for labeling and re-observing locations, as well as for unambiguously communicating
them to other team members. RFIDs can already be located at the disaster site, e.g. they
may be attached to consumer products, or intentionally integrated into building structures.
Alternatively, they can be deployed by rescue teams [Miller et al., 2006, Kleiner
et al., 2006c], either by manually fixing them to relevant locations, or by automatically
deploying them in masses from aerial vehicles. Additionally, RFID tags can be resistant
under extremely hostile conditions. For example, Surface Acoustic Wave (SAW)
technology allows to produce RFID transponders that are known to operate stable up to
several hundreds of degrees, even up to 1000 ◦ C [Fachberger et al., 2006]. They can be
attached to metal surfaces, and are operational within a range of 10 meters for identification,
distance measurements, and temperature measurements (up to 400 ◦ C) [Reindl
et al., 2001,Research, 2007]. Hence, they can be deployed even under harsh conditions,
e.g. in buildings set on fire.
Furthermore, the memory of RFIDs can be utilized by rescue teams for leaving behind
information, such as nearby locations of victims, and nearby explored locations,
supporting the search of other rescue teams in the field. The concept of labeling locations
with information crucial for the rescue task has been already applied in real
disaster response situations. During the disaster relief in New Orleans in 2005, rescue
task forces marked buildings with information concerning, for example, hazardous materials
or victims inside the buildings. A description of these kinds of markings can be
found in a document published by the U.S Dep. of Homeland Security [FEMA, 2003].
2.5 Rescue Robots Freiburg
The work proposed in this thesis was extensively tested on two different robot platforms,
which are a 4WD (four wheel drive) differentially steered robot for the autonomous
team exploration of large office-like areas, and a tracked robot for mapping and overcoming
three-dimensional obstacles, such as stairs and ramps. Both robot platforms
were custom made at the University of Freiburg and designed under the following criteria
(according to Section 2.1):
• Capability of autonomous operation.
• Lightweight design in order to gain a maximal degree of mobility.
• Manufacturing costs.
• Heterogeneous team of robots with different capabilities.

32 Chapter 2. Rescue Robotics Hardware
Figure 2.10: Robots of the team Rescue Robots Freiburg. Left column, top to bottom:
The Zerg robot equipped with a Sick S300 LRF, and navigating in the yellow arena
during the Autonomy Competition at RoboCupRescue 2005 in Osaka. Right column,
top to bottom: The Lurker robot climbing up a ramp during the RescueCamp 2006 in
Rome, and searching for victims in the red arena at RoboCupRescue 2005. A team of
robots waiting for mission start during RoboCupRescue 2005.
Both robot platforms process data from sensors with a CardS12 module, which consists
of a MC9S12D64 Micro Controller Unit (MCU) with a 16-bit HCS12 CPU, 64 KB of

2.5. Rescue Robots Freiburg 33
flash memory, 4 KB RAM, and 1 KB EEPROM. The module offers a large amount of
peripheral connectivity, such as Analog-Digital Converters (ADCs), Pulse Width Modulation
(PWM) ports, and digital I/Os. Furthermore, each robot carries a lightweight
laptop which connects to the MCU via RS-232. The laptop processes the sensor data
further and transmits it, if required, via wireless LAN to an operator station. Details of
the mechanical design and schematics of the electronic components of both robots are
found in the Appendix (Chapter 10).
2.5.1 Lurker robot
Figure 2.10 (right column) shows the tracked Lurker robot, which is based on the Tarantula
R/C toy. Although based on a toy, this robot is capable of climbing difficult obstacles
such as stairs, ramps, and random stepfields. This is possible due to its tracks,
which can operate independently on each side, and the “Flippers” (i.e. the four arms of
the robot), which can be freely rotated at 360 ◦ . The robot is controlled by relays one
for each track (left-hand and right-hand side), and one for each flipper rotation (front
and rear axis), which are connected with four digital outputs of the MCU. The base
was heavily modified in order to enable autonomous operation. First, we added a 360 ◦
freely turnable potentiometer to each of the two axes for measuring the angular position
of the flippers. Second, we added ten touch sensors to each flipper, allowing the
robot to measure force when touching the ground or an object. The data from the ten
touch sensors is multiplexed onto four wires that are bridged with sliding contacts at
the rotating joints. The touch sensors, although simply constructed by push buttons,
provide valuable feedback if the robot touches the ground or an obstacle. This information
together with the angles of the flippers serves as a basis for autonomous skills, and
also supports tele-operation if the information is visualized on the operator interface.
Furthermore, the robot is equipped with a 3-DOF Inertial Measurement Unit (IMU)
from Xsens, allowing drift-free measurements of the three Euler angles yaw, roll, and
pitch, and two Hokuyo URG-X004 Laser Range Finders (LRFs), one for scan matching
(shown in Section 3.2), and one for elevation mapping (shown in Section 5.3), where
the latter can be tilted in the pitch angle within 90 ◦ . For feature tracking, which will
be elaborated on in Section 3.4, a Logitech QuickCam Pro 4000 web cam [Logitech,
2006] has been utilized. The Tarantula toy later on has also been used by other groups
for education and research [Sheh, 2006].
2.5.2 Zerg robot
Figure 2.10 (left column) shows the Zerg robot, a 4WD differentially steered platform,
which was completely hand-crafted. The 4WD drive provides more power to the robot
and therefore allows to drive up ramps and to operate on rough terrain (as far as possible
with the utilized wheel-base). Each wheel is driven by a Pitman GM9434K332 motor

34 Chapter 2. Rescue Robotics Hardware
with a 19.7:1 gear ratio and a shaft encoder. The redundancy given by four encoders
allows to detect heavy slippage and situations in which the robot gets stuck, as will be
shown in Section 3.3. In order to reduce the large odometry error that naturally arises
from a four-wheeled platform, we also utilized an Inertial Measurement Unit (IMU)
from Xsens. Moreover, the robot is equipped with a Thermal-Eye infrared thermo camera
for victim detection. Localization and mapping is performed by a Hokuyo URG-
X004 LRF (indoor version), or a Sick S300 (outdoor version).
2.5.3 Human-Robot Interaction (HRI)
One important goal of Rescue Robotics is to provide technology for assisting the tasks
of humans during urban search and rescue. Human-Robot Interaction (HRI) is an active
research field dealing with the problem of increasing the efficiency of controlling
partially autonomous robot teams, while decreasing the cognitive load of the operator.
Wang et al. demonstrated empirically the advantage of mixed initiative teams, e.g.
partially autonomous robots controlled by a human operator [Wang and Lewis, 2007].
Caspar and Murphy studied human robot-interactions during The World Trade Center
rescue response [Casper and Murphy, 2003]. They made eleven recommendations for
the rescue field based on their findings from robotics, computer science, engineering,
and psychology. Murphy provides an overview on how robots are currently used in
urban search and rescue and discusses the HRI issues encountered over the past eight
years [Murphy, 2004].
Inspired from HRI research recommendations, we developed two applications for allowing
a single operator to control a heterogeneous team of Zerg and Lurker robots.
We specifically designed a Graphical User Interface (GUI), which can be used to fully
control any robot of the team (see Figure 2.11 (a)). The GUI is realized by a similar
approach as proposed by the RoBrno team at RoboCup 2003 [Zalud, 2004]. Images
from video cameras are shown in full size on the screen, and additional information
is overplayed via a Head Up Display (HUD) in order to focus the operator’s attention
on crucial information. The kind of additional information that is displayed, and its
transparency (alpha value), can be adjusted by keyboard commands. Robot control is
carried out with a joypad that is connected to a portable Laptop. Besides images from
the thermo camera and video camera mounted on the robot, the operator receives readings
from other sensors, such as range readings from the LRF, compass measurements,
and the battery state of the robot. Data from the LRF is used to simplify the operator’s
task of navigation, i.e. to estimate the distance to surrounding obstacles. The GUI facilitates
to dynamically switch control between multiple robots that are actively sending
data from the field at the same time.
Furthermore, we developed the strategic tool IncidentCommander for issuing highlevel
commands, e.g. to assign robots to target positions on the map, or to assign complex
behaviors, such as stair climbing tasks, which then are autonomously executed by

2.5. Rescue Robots Freiburg 35
(a)
(b)
Figure 2.11: Human Robot Interaction (HRI): (a) The RoboGui, a graphical user interface
for controlling and monitoring robots, visualizing images from IR and video
camera, and readings from the LRF and IMU. (b) Joypad for operator control.
the commanded unit (see Figure 2.12). The IncidentCommander receives local maps
Figure 2.12: The IncidentCommander, a user interface for map-merging and assigning
high-level commands to multiple robots, visualizing the map generated by the robots,
and images from detected victims.

36 Chapter 2. Rescue Robotics Hardware
generated by the robots and allows to merge them into a global map, either manually
or automatically, based on detected features, such as corners and corridors. Victims
that are detected by the robots are automatically reported to the IncidentCommander
by sending the victim’s location on the map, as well as the according video and IR
images taken at the victim’s location. In Figure 2.12 the victim appears as a red cross
augmented with a number on the map. If the victim is selected with the input device,
all relevant images are shown. This allows for the operator to verify the victim detection
visually, and to erase the marking on the map if the detection was a false-positive.
Furthermore, the map can be modified manually if errors are detected.

3 Pose Tracking in Disaster Areas
3.1 Introduction
To incrementally build a map from scratch and to localize within this map requires good
estimates of local pose displacements during locomotion. Pose tracking is the process
of continuously tracking the pose of a moving object based on sensor readings. For each
discrete time interval ∆t, the pose tracker estimates from the last n sensor observations
o t , o t−1 , o t−2 , · · · , o t−n the local displacement denoted by distance d ∆t and angle θ ∆t .
Pose displacements can be detected by an odometry sensor, e.g. measured by shaft
encoders mounted on the robot’s wheels, by a vision system, e.g measuring optical flow,
or by scan matching algorithms applied to LRF (Laser Range Finder) measurements.
However, wheel odometry becomes arbitrarily inaccurate if the robot has to drive on
slippery ground or even has to climb over obstacles. Particularly on skid-steered platforms,
e.g. tracked vehicles, odometry leads to large measurement errors in rotation.
Vision-based systems depend on a constant flow of features, such as dark corner points,
by which detection might be affected due to harsh visibility conditions, as for example
smoke, dust, and fire. Scan matching techniques can only be applied reliably if the
scanner continuously captures features, such as lines from corners and walls. They require
polygonal structures as typically found in office-like environments [Gutmann and
Schlegel, 1996, Grisetti et al., 2002], which are not necessarily present in unstructured
environments.
As it is important for robots to track their pose for mapping and localization in rescue
scenarios, it is even more important for human beings, e.g. first responders, to know
their location in order to find and rescue victims. Pedestrian Dead Reckoning (PDR)
stands for methods that estimate incrementally the position of a walking person, as
for example in the context of Location Based Services (LBS) [Lechner et al., 2001],
navigation for the Blind [Ladetto and Merminod, 2002b], and emergency responder
tracking. PDR is typically utilized in urban areas where Global Navigation Satellite
System (GNSS) fails due to signal damping or reflection caused by building structures.
In urban environments, GNSS positioning is affected by the so-called multipath propagation
problem [Grewal et al., 2001]. Buildings in the vicinity of the receiver can easily
reflect GNSS signals, resulting in secondary path propagations with longer propagation
time, which distort the amplitude and phase of the primary signal.
In this chapter, approaches for solving the pose tracking problem on robots and on
37

38 Chapter 3. Pose Tracking In Disaster Areas
humans are introduced. They serve as a basis for RFID technology-based SLAM, which
will be described in Chapter 4. Pose tracking on humans is computed from acceleration
patterns measured by body-worn sensors. We adopted a method from Ladetto and colleagues
[Ladetto et al., 2000] that recognizes human footsteps analytically from acceleration
patterns. For this purpose, we utilized an Inertial Measurement Unit (IMU) that
includes tri-axial accelerometers, gyroscopes, and magnetometers ∗ . Utilizing IMUs has
the advantage that the small and lightweight sensors can be attached to humans in an ergonomic
way. They might be integrated into a helmet [Beauregard, 2006], a belt [Amft
et al., 2004], firefighter jacket [Kleiner et al., 2006a], or even into shoes [Foxlin, 2005].
For pose tracking on robots, two novel approaches, one for wheeled robots and one
for tracked robots, are contributed. Since wheel odometry becomes arbitrarily inaccurate
if robots navigate on slippery ground or have to overcome smaller obstacles, a
method for slippage-sensitive odometry has been developed. The introduced method,
which is designed for 4WD robot platforms with over-constrained odometry, infers slippage
of the wheels from differences in the measured wheel velocities. Inference is carried
out by a decision tree that has been trained from labeled odometry data. Furthermore,
we solve the problem on tracked robots by utilizing a consumer-quality camera
for measuring translations, and an IMU for measuring rotations. The proposed method
tracks salient features with the KLT feature tracker [Tomasi and Kanade, 1991] over
images taken by the camera, and computes from the tracked features the translation of
the robot. Translation tracking is computed from a voting of single feature trackings
that are classified by a tile coding classificator [Sutton and Barto, 1998]. The method is
tailored for a simplified motion model of the robot, requiring a set of discrete velocity
commands.
All methods were extensively evaluated in outdoor environments, as well as in USAR
test arenas designed by the National Institute of Standards and Technology (NIST) [Jacoff
et al., 2001]. Our results show that the proposed methods perform robustly and
efficiently in the utilized benchmark scenarios.
The remainder of this chapter is structured as follows. In Section 3.2 state-of-the-art
techniques for pose tracking are described. Then, two solutions for rescue scenarios are
contributed, which are dead reckoning with slippage sensitive odometry in Section 3.3,
and pose tracking by visual odometry in Section 3.4. In Section 3.5 an existing method
for PDR is discussed, and in Section 3.6 empirical results from indoor and outdoor
experiments of the described approaches are presented. In Section 3.7 an overview on
related work is given, and in Section 3.8 we conclude.
∗ Note that a magnetometer measures the earth magnetic field and hence offers the functionality of a
digital compass.

3.2. Conventional techniques 39
3.2 Conventional techniques
3.2.1 Dead reckoning
Dead reckoning is the process of estimating a global position of a vehicle by continuously
advancing a known position by incorporating bearing, speed, and traveled distance.
For wheeled robots, dead reckoning is a method for determining the movement
of the robot by measuring the revolution of the wheels. This is accomplished by shaft
encoders, which are mounted directly on the wheel axes. They provide the number of
counted ticks, which is proportional to the revolution of the wheel. Given the gear ratio
n, the resolution of the decoder (ticks per revolution) R and the diameter of the wheels
D, one can calculate a conversion factor c for the conversion between counted ticks and
the linear wheel displacement:
c = πD
nR . (3.1)
If N r , N l are the tick counts of the right and the left wheel, respectively, one can compute
the right wheel displacement d r = cN r and left wheel displacement d l = cN l . The overall
traveled distance d and traveled angle α can be calculated from the wheel distances d l
and d r after a simplified model from Borenstein [Borenstein et al., 1996]:
d = (d r + d l )
2
, α = (d r − d l )
, (3.2)
B
where B denotes the distance between both wheels, also known as wheel base.
Figure 3.1: Dead reckoning on a skid-steering 4WD robot driving in a cellar: The
resulting map is unusable for planning and navigation.
Dead reckoning is usually accompanied with measurement errors, for example caused
by wheel slip during locomotion, which do accumulate over time. Figure 3.1 depicts the
result from successively integrating range measurements from a laser range finder with

40 Chapter 3. Pose Tracking In Disaster Areas
respect to the odometry pose of the robot. The resulting map is unusable for planning
and navigation.
For any further processing of dead reckoning information it is essential to maintain
a model of the accumulating errors. The traveled distance d and angle α is modeled by
a Gaussian distribution N(u, Σ u ), where u = (d, α) T and Σ u is a 2 × 2 covariance matrix
expressing dead reckoning errors. Likewise uncertainty of the two-dimensional pose
l = (x, y, θ) T is modeled by a Gaussian distribution N (µ l , Σ l ), where µ l is the mean and
Σ l a 3 × 3 covariance matrix [Maybeck, 1990]. Given this representation, the pose at
time t can be updated from input u t as follows:
⎛
l t = F (l t−1 , u t ) = ⎜⎝
x t−1 + cos(θ t−1 )d t
y t−1 + sin(θ t−1 )d t
θ t−1 + α t
⎞⎟⎠ , (3.3)
Σ lt = ∇F l Σ lt−1 ∇F T l + ∇F u Σ u ∇F T u , (3.4)
where
Σ u =
⎛
∇F l = ⎜⎝
⎛
∇F u = ⎜⎝
( dσ
2
d
0
0 ασ 2 α
1 0 − sin(θ t−1 )d t
0 1 cos(θ t−1 )d t
0 0 1
cos(θ t−1 ) 0
sin(θ t−1 ) 0
0 1
)
, (3.5)
⎞
⎞
⎟⎠ , (3.6)
⎟⎠ . (3.7)
Note that the update shown in Equation 3.3 is an approximation and only applicable if
the robot moves almost a straight-line within the time interval in which d and alpha are
measured. The result from incrementally applying the update on a skid-steering robot
driving in a cellar is depicted in Figure 3.1. One method to improve pose estimates
from wheel odometry, is to utilize an Inertial Measurement Unit (IMU) as described in
Chapter 2. Particularly in case of skid-steered vehicles, such as a 4WD robot or tracked
robot, the usage of an IMU is recommendable. Pose tracking from odometry data can
be further improved by utilizing range measurements from the LRF by incremental scan
matching.
3.2.2 Scan matching
The basic idea behind scan matching is to rotate and translate a scan in order to match a
previously taken scan and thereby estimate the displacement (∆x, ∆y, ∆θ) of the robot.
It has been shown that particularly in environments with vertical structures, such as

3.3. Slippage sensitive wheel odometry 41
walls around an indoor soccer field, this can efficiently be performed [Gutmann et al.,
2001]. Various methods were introduced by researches in the past which can generally
be separated into scan point-based and grid-based methods. Among the scan pointbased
methods, Cox and colleges proposed a method particularly suited for polygonal
environments [Cox, 1991]. Their method matches range readings with a priori given
line mode. Lu and Milios proposed a method that can also be applied in non-polygonal
environments [Lu and Milios, 1994]. Gutmann showed how to combine these two
methods in order to improve their overall performance [Gutmann, 2000]. Grid-based
methods have the advantage that they are able to filter-out erroneous scans by averaging
them by an occupancy grid [Moravec and Elfes, 1985], however, their disadvantage is
the high requirement of memory and computational resources. In this section, a gridbased
scan matcher introduced by Hähnel [Hähnel, 2005] will be described in more
detail since this method was implemented within the system described further on.
The technique determines from a sequence of previous scan observations z t , z t−1 , ..., z t−n
subsequently for each time point t an estimate of the robot’s pose k t . This is carried out
by incrementally building a local grid map from the n most recent scans and estimating
the new pose k t of the robot by maximizing the likelihood of the scan alignment of the
latest scan z t in the current map. The robot pose N ( )
l t , Σ lt is fused with the pose of the
scan matcher N ( )
k t , Σ kt by:
l t+1 = ( )
Σ −1 + Σ −1 −1 ( )
k t Σ
−1
l t + Σ −1
k t
k t (3.8)
l t
Σ lt+1 = ( Σ −1
l t
l t
+ Σ −1
k t
) −1
(3.9)
The result from applying scan matching incrementally on laser range data collected by
a robot driving in a cellar is depicted in Figure 3.2. Scan matching-based pose tracking
leads to good results if the scanner captures sufficient features in the environment, such
as walls and corners. However, if the LRF has a restricted field of view, e.g. a range
limit of four meters as the LRF from Hokuyo, scanning leads to a large number of far
readings. Far readings are measurements at maximum range leading to an insufficient
amount of features. This problem can be solved by disabling scan matching and to
continue pose tracking from odometry data only. However, to detect such situations is
difficult since an increase of far readings does not reliably indicate the true amount of
information provided by the scan.
3.3 Slippage sensitive wheel odometry
If the robot operates on varying ground, as for example concrete sporadically covered
with newspapers, or if the robot gets stuck on obstacles, odometry errors are not linearly
growing anymore, but are dependent on the particular situation. Therefore, we designed

42 Chapter 3. Pose Tracking In Disaster Areas
(a)
(b)
Figure 3.2: Scan matching with dead reckoning on a skid-steering 4WD robot driving
in a cellar: (a) The first part of the resulting map is usable for planning and navigation.
(b) As the robot proceeds, the map gets unusable due to constantly increasing
positioning errors.
the Zerg robot with an over-constrained odometry for the detection of slippage of the
wheels by utilizing four shaft-encoders, one for each wheel. From these four encoders,
we recorded data while the robot was driving on varying ground, and labeled the data
sets with the classes C = (slippage, normal). This data was taken to learn a decision
tree [Quinlan, 2003] with the inputs I = (∆v Le f t , ∆v Right , ∆v Front , ∆v Rear ), representing
the velocity differences of the four wheels, respectively. For example, ∆v Front denotes
the velocity difference of the two front wheels, and ∆v Right the velocity difference of
the two wheels on the right hand side of the robot. Under normal operation, ∆v Le f t
and ∆v Right should be close to zero, however under slippage, these differences increases.
As depicted in Figure 3.3, the trained classifier reliably detects this slippage from the
velocity differences.
Given the detection of slippage, the traveled distance d is computed from the minimum
wheel velocity given by v t = min ( )
v Le f tFront , v RightFront , v Le f tRear , v RightRear , and the
robot’s pose is updated according to Equation 3.3, however, with σ 2 d slip
, within covariance
matrix Σ u , in order to increase uncertainty in translation. Note that the rotation
update needs not to be modified since the traveled angle α is measured by the IMU,
which is not affected by wheel slippage. The values for σ 2 d and σ2 d slip
have been determined
experimentally. During extensive runs with slippage events, we recorded the
true traveled distance, determined with scan matching, and the distance estimated by
the odometry. The data set was labeled by the slippage detection and then was utilized
for computing the Root Mean Square (RMS) error for determining the variances σ 2 d and
σ 2 d slip
. We finally determined σ d = 0.816 cm and σ m d slip
= 24.72 cm . As will be shown in
m
Section 3.6, the improved odometry reduces the error significantly, while maintaining
appropriate covariance bounds.

3.4. Visual odometry based pose tracking 43
1000
800
600
Left front
Right front
Left rear
Right rear
Speed [mm/s]
400
200
0
-200
-400
-600
-800
0
Forward Turn Forward Slippage Turn Slippage Turn
2 4 6 8 10 12 14 16 18
Time [s]
Slip detected
True
False
0
2
4
6
8
10
Time [s]
12
14
16
18
Figure 3.3: Slip detection on a 4WD robot: each line in the upper graph corresponds
to the velocity measurement of one of the four wheels by shaft encoders. The black
arrows indicate the true situation, e.g. driving forward, slippage, etc., and the red line in
the lower graph depicts the automatic slip detection by the decision tree classifier, given
the velocities as input.
3.4 Visual odometry based pose tracking
In this section a solution to pose tracking on tracked robots that can support SLAM during
autonomous behaviors on three-dimensional obstacles is described. The introduced
approach aims at an inexpensive, light-weight, and computational efficient implementation,
as for example based on a single consumer-quality webcam. Since the robot is
equipped with an IMU providing high accuracy orientation information, and the motion
model on tracked robots can be simplified by restricting it to constant velocities,
our goal is to distinguish reliably between forward, backward, and none translations,
only. This limitation is motivated by the fact that computing the full metric information
of the translation, e.g. to determine the scale, requires to extract depth information from
camera observations, which typically can only be done by stereo vision. However, the
problem can be relaxed if robots operate in planar environments [Weigel et al., 2002],
or by utilizing metrical initialization points with known spacial displacement in the environment
[Davison, 2003]. By focusing on determining the translation direction only,

44 Chapter 3. Pose Tracking In Disaster Areas
we solve this problem for three-dimensional environments as the robot’s known driving
speed produces a reference that can be used to generate metrical pose information.
3.4.1 Feature tracking
Salient features are tracked over multiple images with the KLT feature tracker [Tomasi
and Kanade, 1991], and utilized for calculating difference vectors that indicate the
robot’s motion. Since rotations are estimated by the IMU sensor, and the goal is to
determine translations from the images, the camera is mounted to either the left or right
side of the robot, yielding the effect that forward and backward motion results in horizontal
tracking vectors pointing into the direction of the motion. The true translation
of the robot is determined based on the individual voting of single translation vectors.
Each vector votes for one of the possible translations according to a pre-trained tile
coding classificator [Sutton and Barto, 1998].
In general, an image sequence can be described by a discrete valued function I(x, y, t),
where x, y describe the pixel position and t describes the time. It is assumed that features
detected in an image also appear in the subsequent image, however translated by d =
(ξ, η) T , where ξ and η denote the translation in x and y:
I(x, y, t + τ) = I(x − ξ, y − η, t) (3.10)
Usually, a feature tracker determines this translation by minimizing the squared error ɛ
over a tracking window. For brevity we define I(x, y, t + τ) as J(x) and I(x − ξ, y − η, t)
as I(x − d), leading to the following error measure with a weighting function w [Tomasi
and Kanade, 1991]:
∫
ɛ = [I(x − d) − J(x)] 2 wdx. (3.11)
W
To facilitate the process of feature tracking, the selection of appropriate features, i.e.
features that can easily be distinguished from noise, is necessary. Hence, features that
show light-dark changes, e.g. edges, corners, and crossings, are selected with high probability
by the KLT feature tracker. In Figure 3.4, examples of KLT’s adaptive feature
selection and the tracking over a series of images are shown.
When traversing obstacles, the robot’s motion is not exclusively a forward or backward
motion. Instead, it is overlaid with noise that originates from slippage of the tracks
and shaking of the robot’s body due to rough terrain, leading to jitter effects. Since these
effects usually do not accumulate over time, our method generates trackings over multiple
frames, rather than performing single frame trackings only.
If trackings of the same feature coexist over more than two images, their corresponding
translation vectors d i , d i+1 , ..., d k are replaced by a single translation vector d ik , consisting
of the vector sum of all trackings between d i and d k . In order to reward trackings
over multiple frames, a weight w ik = |k − i| is assigned to each tracking. These weights

3.4. Visual odometry based pose tracking 45
(a)
(b)
(c)
(d)
Figure 3.4: KLT feature tracking: (a,b) Features (red dots) are adaptively selected
within images. (c) Feature tracking over two subsequent images. The vectors between
two corresponding features, shown by red lines, indicate the movement of the camera.
(d) The tracking over a series of five images.
are used during the voting process, which will be described in Section 3.4.3.
3.4.2 Filtering of rotations around the pitch angle
Since the focus lies on translation estimation, rotations have to be filtered-out in advance.
It is assumed that translations are free of rotations around the yaw angle since
otherwise the motion is considered as a rotation. However, during straight-line navigation,
rotations might occur around the pitch angle while climbing over obstacles and
movements of the flippers that change the robots pitch with respect to the ground. Due
to the high variance in the latency time of the employed camera system (a web cam connected
via USB 1.1), this can only be accurately achieved on the image data directly,
rather than by combining image data with rotation angles from the IMU sensor. Given
a feature tracking between two images of the form (x i , y i ) T → (x j , y j ) T , which includes

46 Chapter 3. Pose Tracking In Disaster Areas
a rotation around the point (r x , r y ) T with angle α, one can derive a corresponding rotation
free tracking (x i , y i ) T → (x ′ j , y′ j )T after the following equation, with given rotation
matrix R (.):
(x ′ j, y ′ j) T = (r x , r y ) T + R(−α) · (x j − r x , y j − r y ) T (3.12)
Therefore, in order to perform the filtering of rotations, one has to determine the rotation
center (r x , r y ) T and rotation angle α. Rotating points of different radii describe
concentric circles around the rotation center. When considering two feature trackings
whose features lie on a circle, one can see that the perpendicular bisectors of the two
lines, given by the connecting start- and endpoint of the feature tracking, subtend in the
rotation center, as shown in Figure 3.5 (a).
(a) (b) (c) (d)
Figure 3.5: (a) The perpendicular bisectors (green) of the tracking vectors (red) subtend
at the center of the circle (magenta). (b) Example of the Monte Carlo algorithm:
The perpendicular bisectors (green) point to the center of rotation (magenta). Red dots
depict the sampled intersection points. (c,d) Example of the rotation correction while
the robot changes the angles of its front flippers. The feature vectors before (c) and after
(d) the correction.
This property is exploited with a Monte Carlo algorithm for estimating the true center
of rotation (see Figure 3.5 (b)). First, up to n possible centers of rotation are sampled
from the set of feature trackings T by Algorithm 1. Second, all sampled centers of
rotation are put into a histogram, where the final center is determined by the histogram’s
maximum.
Furthermore, one has to determine the rotation angle, which can be done by calculating
the vector cross product. Given a feature tracking (x i , y i ) T → (x j , y j ) T rotated
around (r x , r y ) T by α, one can calculate the cross product by considering the start- and
endpoint of the feature tracking as endpoints of vectors starting at the rotation center.
Suppose v i = (x i −r x , y i −r y ) T and v j = (x j −r x , y j −r y ) T are vectors derived from tracking
images I and J, respectively. Then, the angle between these vectors α = ∠(v i , v j ) can
be calculated from the vector product: v i × v j = ||v i || · ||v j || · sin(α). Given the rotation
center (r x , r y ) T from the previous estimation, one can determine the true rotation angle α
by averaging rotation angles from all single feature trackings. Finally, it is necessary to
prevent the algorithm from being executed on rotation-free sequences. This is achieved
by adding a center of rotation to the histogram, only if it is located within the bounding

3.4. Visual odometry based pose tracking 47
Algorithm 1: Sample up to n possible centers of rotation
Input: A set of feature trackings: T
Output: A set of calculated intersection points: C
C = ∅;
for i = 0; i < n; i++ do
t 1 ← selectRandomFeatureTracking(T);
t 2 ← selectRandomFeatureTracking(T);
s 1 ← calculatePerpendicularBisector(t 1 );
s 2 ← calculatePerpendicularBisector(t 2 );
(cut, det) ← calculateIntersectionPoint(s 1 , s 2 );
if det < minDeterminant then
continue;
end
C ← C ∪ cut;
end
box of the image. Rotation centers that are far from the bounding box are most likely
due to quasi-parallel feature translations, which in turn indicate a rotation-free movement.
If the number of centers of rotation is below a threshold λ, the transformation
of Equation 3.12 is not applied. We determined experimentally λ = 10 [Dornhege and
Kleiner, 2006].
3.4.3 Classification
From the set of filtered translation vectors, one can determine the robot’s translation.
However, the translation vectors are given in the two dimensional image plane. The
projection from translation vectors of the vision system to the robot’s translation depends
on the intrinsic parameters of the camera, e.g. focal length and lens distortion,
and on the extrinsic parameters of the camera, e.g. the translation and rotation relative
to the robot’s center. This projection can either be determined analytically or by a mapping
function. Due to the assumption of a simplified kinematic model, this mapping
can be learned efficiently by a function approximator, classifying each vector into one
of the classes C = { f orward, backward, none}.
The learning is based on collected data which was automatically labeled during teleoperation
runs under mild conditions, i.e. without heavy slippage. During a second
phase, the data labeling has been verified manually by a human on a frame to frame
basis. This procedure allows for the efficient labeling of thousands of trackings since
single images contain several features. Each labeled tracking is described by the class
assignment c ∈ C and the vector v = (x, y, l, α) T , where x, y denotes the origin in the
image, l denotes the vector length and α denotes the vector heading. Given the labeled

48 Chapter 3. Pose Tracking In Disaster Areas
data, tile coding function approximation is used for learning the probability distribution
P(c | x, y, l, α). (3.13)
Tile coding is based on tilings which discretize the input space in each dimension.
Shape and granularity of these discretizations can be adjusted according to the task.
Multiple tilings are overlaid with a randomized offset in order to facilitate generalization.
During learning, each tile is updated according to: w i+1 = w i +α · (p i+1 −w i ), where
w i is the weight stored in the tile, p i ∈ {0, 1} is the manually labeled class assignment,
and α is the learning rate, which is set to 1 , where m is the number of overlaid tilings in
m
order to ensure normalized probabilities [Sutton and Barto, 1998]. Based on the probability
distribution induced by P(c | x, y, l, α) over C, each vector v i votes individually
for a class assignment c i with respect to its location, length, and heading:
c i = argmax P(c | x i , y i , l i , α i ) (3.14)
c∈C
Let c k i
= I (c i = k) be the class indicator function, which returns 1 if c i = k and 0
otherwise. Then, the final classification a for each frame can be decided based on the
maximal sum of weighted individual votes from each vector:
a = argmax
k∈C
N∑
c k i · w i (3.15)
i=1
Note that w i increases according to the number of times the underlying feature has been
successfully tracked by the feature tracker previously described.
In order to determine the distance d traveled between two images I and J, a constant
translational velocity v T of the robot † is assumed. Given time stamp t j and t i of image
I and J, respectively, d can be calculated by:
⎧
v T · (t j − t i ) if class = forward
⎪⎨
d = −v T · (t j − t i ) if class = backward
(3.16)
⎪⎩ 0 otherwise
Finally, from the yaw angle θ of the IMU and the robot’s last pose (x old , y old , θ old ) T the
new pose of the robot is calculated by:
(x new , y new , θ new ) T = (x old + d · cos(θ old ), y old + d · sin(θ old ), θ old ) T (3.17)
† Note that this value could also be automatically adjusted according to the vehicles current set-velocity.

3.5. Pedestrian Dead Reckoning (PDR) 49
3.5 Pedestrian Dead Reckoning (PDR)
In this section a robust method for estimating the distance displacement d and azimuth
angle (heading) θ of walking persons, by utilizing a single IMU sensor that contains
a tri-axial accelerometer, a tri-axial gyroscope, and a tri-axial magnetometer, will be
described. According to Section 3.2.1, the displacement vector u = (d, θ) T with covariance
matrix Σ u is further processed by Kalman-based dead reckoning, yielding the pose
estimate ˆl = ( ˆx, ŷ, ˆθ) T with covariance matrix Σ l .
3.5.1 Distance estimation
Generally, the traveled distance d of a walking person can only under difficulties be
computed by double integration of the antero-posterior (forward-backward) acceleration
[Ladetto et al., 2000, Foxlin, 2005]. This has two reasons: first, the alignment of
the IMU device on the body of the person strongly influences the measurements, hence
it has to be calibrated each time the position of the device has been changed. Second,
the double integration quickly accumulates measurement errors, e.g. caused by abrupt
acceleration changes during walking, leading to non-permissible positioning errors. An
alternative approach is to detect the step occurrence from vertical and horizontal acceleration
patterns merged with a physiological model of human walking. Experiments
have shown that the maximal step frequency of humans is about 5 Hz, and that there
are no acceleration peaks beyond 15 Hz. Therefore, according to the law of Shannon,
accelerations of walking patterns can be reliably measured at a sample frequency
of 30 Hz. There are several identification strategies, ranging from computational expensive
Fourier analysis to simple Zero-crossing, which basically detects a step if the
acceleration curve crosses the zero value in an ascending way. A method that shows
very robust results within the least computation time has been introduced by Ladetto
et al. [Ladetto et al., 2000]. Human walking generates a vertical acceleration with a
maximum value if a foot is placed on the ground (see Figure 3.6). By detecting these
maxima in the vertical acceleration curve within a fixed time interval, it is possible to
detect and count the occurrence of foot steps. The method can be separated into three
parts, which are step detection, step counting, and direction detection.
To detect the walking mode of the person, e.g. to distinguish between forward walking,
lateral (side-wards) walking, and standing, is a challenging task due to the individual
walking style of humans. Ladetto concluded that the mean variance over a short
period of the vertical (up-down), and antero-posterior (forward-backward) acceleration
correlates significantly with the walking mode and step frequency. The higher the frequency,
the higher the variance. The occurrence of steps can thus be determined from
the mean variance computed within a fixed time interval. Furthermore, a comparison
between the mean variance of the antero-posterior acceleration and vertical acceleration
allows to distinguish between lateral (side-wards) and forward movements. A lateral

50 Chapter 3. Pose Tracking In Disaster Areas
Figure 3.6: Vertical (up-down), and antero-posterior (forward-backward) acceleration
patterns recorded during human walking. Courtesy of Quentin Ladetto.
movement is detected, if the variance of the vertical acceleration is higher than the one
of the antero-posterior, and vice versa.
To detect the frequency of steps, one has basically to count the maxima in the anteroposterior
acceleration curve. Depending on the individual walking style of a person,
a step might cause two peaks located closely to each other, originating from the impacts
of the heel and the sole with the ground, respectively, whereas the heel impact
normally shows the bigger amplitude. In order to detect step maxima reliably, it is
beneficial to filter-out signal disturbances in advance. This is achieved by applying a
strong low-pass filter that removes high frequency components from the signal, e.g.
frequencies above 5 Hz. Given the filtered signal with each local maxima representing
a footstep, the step frequency is determined by time differencing the maxima. The
traveled distance is computed by multiplying the frequency with the individual length
of the person’s steps. Since the step length greatly varies among different persons, we
automatically calibrate this parameter from distances computed by GNSS positions if
they are available [Ladetto et al., 2000].
Furthermore, the direction of walking (forward-backward, left-right) is determined
by computing the angles between the maxima and the two zero-crossings of the signal
before and after the maxima, respectively. For example, a backward motion is detected
if the angle between the maxima and the previous zero-crossing is bigger than
the angle between the maxima and the following zero-crossing. A more detailed description
of the method is found in the work of Ladetto and colleagues [Ladetto et al.,
2000,Gabaglio, 2003,Ladetto and Merminod, 2002b]. For the experiments described in

3.6. Experimental results 51
this thesis an implementation of Ladetto’s work from Michael Dippold [Dippold, 2006]
has been used.
3.5.2 Heading estimation
The compass and the gyroscope of the IMU are used in conjunction in order to compensate
perturbations from local magnetic fields, as for example caused by ferromagnetic
building materials such as reinforced concrete or steel [Ladetto and Merminod,
2002a, Ladetto et al., 2001]. This approach ideally combines the strength of both sensors
and compensates their weaknesses at the same time. Whereas the compass provides
absolute measurements of the azimuth, which are partially inaccurate due to local
magnetic fields, the gyroscope measures, also under magnetic influence, the angular
acceleration with high accuracy. However, long term measurements of the gyroscope
are affected by an inherent drift error. Hence, drift errors of the gyroscope can be compensated
by globally accurate azimuth angles from the compass, and local disturbances
of the compass can be compensated with locally accurate measurements of the angular
acceleration.
For the experiments presented in this thesis, an MTi IMU from Xsens was used (see
Chapter 2, Section 2.2.1). The sensor unit provides a special modus called AMD (Adapt
to Magnetic Disturbances) which can be activated for environments containing heavy
magnetic fields. The device was attached to a test person for measuring the vertical and
lateral acceleration, as well as the azimuth orientation (see the orange box in Figure 3.7).
Based on an empirical evaluation, we modeled pose tracking uncertainty with σ 2ˆd =
(0.05 m) 2 d, and σ 2ˆθ = (15◦ ) 2 .
3.6 Experimental results
In this section, results from both simulated and real-robot experiments are provided. All
real-robot experiments were carried out on the robot platforms described in Chapter 2
in outdoor scenarios, and testing arenas that are equal or similar to those proposed by
NIST. Experiments on pedestrian dead reckoning were carried out with the MTx IMU
from Xsens and a SIRFstarIII chip-based GPS device. Results from pose tracking on
robots, specifically, wheeled pose tracking and visual odometry-based pose tracking, are
presented in Sections 3.6.1 and 3.6.2, respectively, whereas results from dead reckoning
on pedestrians are presented in Section 3.6.3.
3.6.1 Pose tracking under heavy slippage
The slippage detection method has been extensively evaluated on the Zerg robot. During
this experiment, the robot performed different maneuvers, such as moving straight,

52 Chapter 3. Pose Tracking In Disaster Areas
Figure 3.7: Test person with Xsens MTi and Holux GPS device for obtaining ground
truth data.
turning, and accelerating while driving first on normal and then on slippery ground.
Afterwards, each situation has been manually labeled with one of the six classes slipstraight,
slip-turn, slip-accelerate, noslip-straight, noslip-turn, and noslip-accelerate.
Table 3.1 summarizes the results of the classification, where bold numbers indicate the
correct classification, i.e. true-positives. The method is able to reliably detect slippage
even while the robot is accelerating or performing turns.
Classification
❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤
True situation
Slip No Slip
Straight
No Slip 10 (0.5%) 2051 (99.5%)
Slip 2363 (90.1%) 236 (8.9%)
Turn
No Slip 28 (0.9%) 3226 (99.1&)
Slip 2684 (96.4%) 102 (3.6%)
(De-)Acceleration
No Slip 75 (14.9%) 426 (85.1%)
Slip 126 (98.5%) 2 (1.5%)
Table 3.1: Classification accuracy of the slippage detection under different maneuvers
of the robot. Bold numbers indicate the correct classifications, i.e. true-positives.
In order to evaluate the slippage detection-based improvement of the odometry, we
conducted experiments for the comparison of both improved and conventional odometry
and their covariance bounds. Figure 3.8 shows the performance of slippage sensitive
odometry compared to conventional odometry. It can be seen in Figure 3.8 (a) that the

3.6. Experimental results 53
error of the conventional odometry increases drastically during slippage (taking place
between 10 and 20 meters). Moreover, the covariance bound significantly underestimates
the error. However, in the same situation, slippage sensitive odometry is capable
of reducing the error (Figure 3.8 (b)), while providing valid covariance bounds.
10000
9000
Odometry error
3σ Bound
Conventional odometry
10000
Odometry error
3σ Bound
Slippage sensitive odometry
8000
8000
7000
Error [mm]
6000
5000
4000
Error [mm]
6000
4000
3000
2000
2000
1000
0
0
5
10
15
Distance [m]
(a)
20
25
30
0
0
5
10
15
Distance [m]
(b)
20
25
30
Figure 3.8: Conventional odometry (a) compared to slippage sensitive odometry (b)
during the event of slippage (between 10 and 20 meters): In contrast to conventional
odometry, improved odometry reduces the position error (blue line) and provides valid
covariance bounds (red line) during slippage.
(a)
(b)
Figure 3.9: Zerg robot during the final of the Best in Class autonomy competition at
RoboCupRescue 2005 in Osaka: (a) slipping on newspapers and (b) the autonomously
generated map. Red crosses mark locations of victims which have been found by the
robot.
The approach of slipping detection has been utilized during the RoboCup Rescue

54 Chapter 3. Pose Tracking In Disaster Areas
competition. Figure 3.9 depicts the Zerg robot during the final of the “Best in Class Autonomy”
competition, held in the NIST arena for Urban Search and Rescue (USAR) [Jacoff
et al., 2001] during RoboCup 2005. In this scenario, robots had to explore an unknown
area within 20 minutes autonomously, to detect all victims, and finally to deliver
a map sufficient for human teams to locate and rescue the victims. Conditions for exploration
and SLAM were intentionally made difficult. For example, the occurrence of
wheel slip was likely due to newspapers and cardboards covering the ground, which
was partially made of steel and concrete. Stone bricks outside the robot’s field of view
caused the robot to get stuck, and walls made of glass caused the laser range finder to
frequently return far readings. We applied computer vision techniques on images generated
by a thermo (IR) camera in order to estimate the relative locations of victims,
if they were detected withing the camera’s FOV. More details on victim detection are
given in Chapter 7. As shown in Figure 3.9, the system was able to cope with these
difficulties and also to build reliable a map augmented with victim locations detected
by the robot. Finally, the system won the autonomy competition in 2005.
3.6.2 Pose tracking on tracked vehicles
The approach of visual odometry was extensively tested on both the tracked robot
Lurker, operating on three-dimensional obstacles, and the wheeled robot Zerg, operating
on flat surfaces. Experiments with the Zerg robot have the advantage that the
visual odometry, and conventional wheel odometry can be compared directly to ground
truth data. They allow a comparison of both methods under the same circumstances and
show, that the visual odometry produces results of comparable accuracy to wheel odometry.
On this robot, position ground truth was determined by LRF-based scan matching,
whereas ground truth on three-dimensional obstacles was measured manually.
Run Trav. dist. [m] Vis. odo. [cm/m] Wh. odo. [cm/m]
lab_1 (2D) 91.53 7.82 ± 1.84 6.17 ± 1.54
lab_2 (2D) 73.72 8.25 ± 2.46 7.59 ± 1.94
cellar (2D) 108.91 19.30 ± 13.65 21.22 ± 12.40
ramp (3D) 6.36 13.28 ± 9.2 -
palette (3D) 2.37 22.08 ± 8.87 -
Table 3.2: Relative error of the visual odometry and conventional wheel odometry
compared to ground truth data (either manually measured for 3D runs or estimated by
scan matching for 2D runs).
Table 3.2 gives an overview on the measured mean and standard deviation of the relative
distance error from visual odometry and wheel odometry on both robot platforms.
Since the method has been mainly developed for tracked vehicles, the Zerg’s kinematic

3.6. Experimental results 55
has been modified in order to be similar to that of the evaluated tracked vehicle, i.e. to
allow only a subset of possible velocities, which are in case of the Lurker robot: stop,
forward, and backward. The results clearly show that on the Zerg platform the visual
odometry reaches an accuracy comparable to the conventional odometry. In the cellar
environment, the visual odometry turned out to be even slightly superior, which can be
explained by the higher degree of wheel slippage that we noticed in this environment.
Comparing the graphs in Figures 3.10 (a) and (b) of the cellar environment with those
of the lab in Figures 3.10 (c) and (d) it becomes clear that the cellar environment is
much harder for both algorithms. The wheel odometry usually suffers from wheel slippage
while the visual odometry has to cope with the fact that the white walls do not
provide good features. Figures 3.10 (a) and (c) depict the accumulation of the distance
error of the wheel odometry and Figures 3.10 (b) and (d) show the distance error of
the visual odometry in the cellar and lab environment, respectively. The real advantage
of the visual odometry, however, is revealed if the robot operates on three-dimensional
obstacles.
The results in Table 3.2 indicate that the introduced method, when applied while
operating on three-dimensional obstacles, provides a usable estimate of the robot’s motion.
Figures 3.11 (a) and (b) depict the driven distance during locomotion over threedimensional
obstacles compared to ground truth distances. The results indicate that, in
case of tracked robots, the tile coding classification and voting applied to a simple kinematic
model lead to sufficiently accurate results. From log files it has been determined
that during the cellar run 87% (96%), the ramp run 81% (93%), and the palette run 94%
(99%) of the classifications detected the correct motion of the robot, where numbers in
brackets denote the voting-based improvement. While processing an image resolution
of 320×240 on a IntelPentiumM, 1.20 GHz, we measured an average processing time of
24.08 ± 0.64 ms for the complete processing without KLT feature tracking. This leads,
together with the feature tracker, to a maximal frame rate of 5.34 ± 1.17 Hz. If processing
an image resolution of 160 × 120, the complete processing without KLT feature
tracking needs 8.68 ± 0.3 ms and allows a total frame rate of 17.27 ± 1.81 Hz. Experiments
proposed in this thesis were carried out with the higher resolution. However,
experiments with the lower resolution showed that these results lead to a comparable
accuracy, too.
The following two experiments demonstrate the application of visual odometry in
the context of SLAM on tracked vehicles negotiating obstacles. We conducted the
experiments on the Lurker robot by utilizing the visual odometry together with the scan
matching algorithm. During the first experiment, the LRF sensor was automatically
controlled by the measured pitch orientation of the robot in order to stay continuously
at a horizontal position. This allows the LRF to perceive the environment independently
from the robot’s orientation, i.e. to return the same laser scan at the same locations also
if the orientation differs. The result is shown by the image series in Figure 3.12 (af).
In (a) and (d) an overview on both obstacles is given, whereas (b) and (e) depict

56 Chapter 3. Pose Tracking In Disaster Areas
45000
40000
Odometry error
3σ Bound
Wheel odometry (Cellar environment)
45000
40000
Odometry error
3σ Bound
Visual odometry (Cellar environment)
35000
35000
30000
30000
Error [mm]
25000
20000
15000
Error [mm]
25000
20000
15000
10000
10000
5000
5000
0
20
40
60
Distance [m]
(a)
80
100
0
20
40
60
Distance [m]
(b)
80
100
45000
40000
Odometry error
3σ Bound
Wheel odometry (Lab environment)
45000
40000
Odometry error
3σ Bound
Visual odometry (Lab environment)
35000
35000
30000
30000
Error [mm]
25000
20000
15000
Error [mm]
25000
20000
15000
10000
10000
5000
5000
0
0
20
40
Distance [m]
(c)
60
0
0
20
40
Distance [m]
(d)
60
Figure 3.10: The accumulating distance error of the visual odometry method compared
to ground truth data: (a,b) measured in a cellar of 15 m × 50 m, (c,d) measured in the
robotic lab, a 5 m × 5 m squared area, (a) and (c) show results from the wheel odometry,
(b) and (d) display the errors originating from visual odometry.
the generated features, and (c) and (f) show the generated maps, at the corresponding
positions, respectively.
As shown by the “black wall” in front of the robot (Figure 3.12 (c)), there are situations
in which the 2D LRF cannot provide sufficient evidence for the robot’s motion.
In this particular case, measurements of the laser are nearly independent of the robot’s
location on the ramp, whereas the visual odometry (see Figure 3.12 (b)) provides clear
motion evidence. Note that the map representation shown in (c) and (f) does not suffice
for the particular task since it does not distinguish between measurements of different
height values at the same location, i.e. parts of the map in (c) have been deleted in
the subsequent map (f). This problem can be solved by utilizing elevation maps, as
discussed in Chapter 5.
In another experiment, the influence of visual odometry on elevation mapping has
been evaluated. Figure 3.13 depicts two elevation maps of the same ramp, one with

3.6. Experimental results 57
350
300
250
Ground Truth
Visual Odometry
Odometry while climbing a ramp
250
200
Ground Truth
Visual Odometry
Odometry while climbing a pallet
Distance [cm]
200
150
100
Distance [cm]
150
100
50
0
50
-50
0 1000 2000 3000 4000 5000 6000 7000
Time [ms]
(a)
0
0 1000 2000 3000 4000 5000 6000 7000
Time
(b)
Figure 3.11: The distance of the visual odometry method compared to manually generated
ground truth data: Results from driving forward and backward on a ramp (a),
and climbing over a wooden palette (b). The red curve (crosses) indicates the manually
measured ground truth, and the blue curve (asterisks) indicates the distance estimated
by visual odometry, respectively.
(a) (b) (c)
(d) (e) (f)
Figure 3.12: Lurker robot performing vision-based pose tracking while overcoming
three-dimensional obstacles: (a,d) the three-dimensional obstacles, (b,e) the features
and classified directions generated from a side camera, and (c,f) the grid maps generated
at these locations, respectively.

58 Chapter 3. Pose Tracking In Disaster Areas
(a)
(b)
1600
1400
Scan Matching error
3σ Bound
Scan matching
1600
1400
Combined error
3σ Bound
Visual odometry with scan matching
Accumulated Error [mm]
1200
1000
800
600
400
Accumulated Error [mm]
1200
1000
800
600
400
200
200
0
0
1
2 3
Distance [m]
(c)
4
5
0
0
1
2 3
Distance [m]
(d)
4
5
Figure 3.13: Comparing elevation mapping based on scan matching only (a) and scan
matching combined with visual odometry (b). The scan matchings small error (c) grows
rapidly out of its usual error bound, when the robot drives on the ramp, while the visual
odometry (d) is not influenced by this effect. Scan matching without visual odometry
support does not correctly reflect the true length of the ramp, because insufficient motion
evidence causes the map to be compressed partially.
support of visual odometry, and one without. The corresponding error graphs show
that scan matching cannot correctly reflect the robot’s motion when the robot drives
on the ramp between 2.75 m and 4.75 m. Usually, estimating the robot’s pose based
on two dimensional scan alignment can be done reliably, but this is no longer possible
when the two-dimensional reference system changes. The result is a rapid error growth
(Figure 3.13 (c)) that leads to distortions in the map due to the incorrect pose assumption.
Fusing distance estimates from the visual odometry, that are not influenced by
this effect, into the scan matchings pose clearly reduces the error. Mapping based on
scan matching only yields a compressed map since in this environment 2D laser scans

3.7. Related work 59
do not provide sufficient information on the motion of the robot, whereas generating a
map based on visual odometry reveals the true size of the ramp, which was verified by
measuring the ramp’s dimensions manually.
3.6.3 Pose tracking of human walking
The PDR implementation described in Section 3.5 has been extensively tested in outdoor
scenarios. Outdoor scenarios have the advantage that ground truth data can be generated
from differential GPS positioning yielding an error within a few meters. Hence,
results from dead reckoning can be compared directly with the true trajectory of the
person.
In order to evaluate the setting, we recorded trajectories of a pedestrian walking for
approximately 3 kilometers around a lake in Freiburg. During these experiments, the
test person was naturally walking at different speeds, either in a forward or side wards
direction, as well as stopping at certain locations. The experiments did not include
unusual motions, such as crouching or climbing. Figure 3.14 shows the resulting trajectories
from different experiments. The red line indicates the ground truth generated
from GPS readings, and the green line depicts the trajectory generated with the PDR
module. As expected, the PDR trajectory diverges from ground truth, the longer the person
walks. Particularly in Figure 3.14 (b), after a walk of approximately 3 kilometers,
the PDR position has nearly an error of 240 meters. In Chapter 4 it will be shown how
these trajectories can be improved by utilizing observed correspondences of RFIDs.
3.7 Related work
Borenstein et al. introduced a method for improving the odometry on differential-drive
robots [Borenstein, 1996]. A method for odometry improvement and optimization of
motor control algorithms on 4WD robots has been introduced by Ojeda et al. [Ojeda
and Borenstein, 2004]. They applied “Expert Rules” in order to infer the occurrence of
wheel slip.
The approach of visual odometry was extensively studied in the past. Davison and
colleagues developed a system for visual EKF-based SLAM using only a single camera
[Davison, 2003]. Their system was mainly evaluated in desk-like environments.
However, in contrast to the system introduced in this chapter, their approach requires
a calibration of the camera at start-up in order to estimate the three-dimensional realworld
locations of the features. Corke and colleagues introduced a solution for a planar
rover equipped with an omni-directional vision system [Corke et al., 2004]. In contrast
to the proposed work, which also aims at indoor applications, they assume that the robot
operates in an open space, as it is usually the case on planetary analog environments.
Milella and Siegwart proposed a system that computes the 6 DOF ego motion of an all-

60 Chapter 3. Pose Tracking In Disaster Areas
(a)
(b)
(c)
(d)
Figure 3.14: Comparison between PDR (green trajectory) with GPS ground truth (red
trajectory). Trajectories have been recorded while walking around a lake in Freiburg
terrain robot based on the Iterative Closest Points (ICP) method processing data from a
3D stereo vision system [Milella and Siegwart, 2006]. Nister and colleagues presented
a system for visual odometry that works with both mono and stereo vision [Nister et al.,
2004]. Their results show that data processing of a stereo system leads to a highly accurate
estimate of the robot’s pose, which was also confirmed by the work of Helmick
and colleagues [Helmick et al., 2004]. The use of a stereo system generally has the
advantage that the depth information of features tracked by the vision system can be
utilized for computing the velocity of the robot. Results proposed in this chapter show
that with the simplified kinematics of tracked robots, a single but lightweight camera
solution can also lead to sufficiently accurate pose estimates.
PDR methods were extensively studied in the past. Human motion has been tracked
by vision sensors [Zhu et al., 2006], as well as based on the analysis of acceleration

3.8. Conclusion 61
patterns [Amft et al., 2004,Foxlin, 2005,Ladetto et al., 2000,Judd, 1997]. Zhu and colleagues
proposed a system that combines data from an IMU sensor and GNSS device
with dead reckoning information generated from a stereo vision package [Zhu et al.,
2006]. However, since their system requires the user to wear two cameras mounted on
a backpack, it cannot be applied in narrow scenarios. Due to the computational power
of small devices and due to the high demand from a wide range of social and military
applications, the development of light-weight solutions for pedestrian dead reckoning
has been significantly accelerated in the past. The approach introduced by Ladetto
and colleagues [Ladetto et al., 2000] has been commercialized and is available as a
hardware implementation, namely the Personal Navigation Module (PNM) [Vectronix,
2007]. The PNM, which is completely integrated into a box that weighs 34 grams and
measures 43 × 34 × 25 mm, was developed by the Ecole Polytechnique Fedalrale de
Lausanne (EPFL) together with the Leica Vectronix AG. It is worth mentioning that the
system won the Swiss Technology Award in 2003. A comparable system is the Dead
Reckoning Module (DRM), which was developed by Tom Judd [Judd, 1997] at the
Point Research Cooperation and the fifth generation (DRM5) is commercially available
from Honeywell [Honeywell, 2007]. The system was evaluated in the context of disaster
response for the localization of first responders [Miller et al., 2006]. During these
experiments persons were tracked with the DRM while walking through buildings of
different types.
3.8 Conclusion
In this chapter methods for pose tracking on tracked and wheeled robots, as well as for
pedestrian walking, were introduced. Methods for tracking robots were mainly evaluated
in the testing arenas proposed by NIST for Urban Search and Rescue. Experimental
results showed that under modest computational requirements good results have been
achieved, even under the harsh environmental conditions found in the arenas of the
NIST benchmark. While pose tracking based on slippage sensitive odometry enabled
mapping during heavy slippage, the visual odometry method allowed to improve pose
estimates significantly while navigating on rough terrain.
As results from PDR experiments showed, the tracking of human beings based on
accelerometers leads to promising results. However, also these results are only preliminary
compared to the situations first responders are exposed to while rescuing victims
after a real disaster. They might crouch, run, and climb within very different and individual
modes, generating a wide range of ambiguous sensor values. Also in this area
further research has to be undertaken in order to capture the whole range of human motion.
Nevertheless, results as they were proposed in this chapter are sufficient for dead
reckoning in scenarios that do not require special modes of motion.

4 RFID Technology-based SLAM
4.1 Introduction
Truly autonomous robot navigation in an unexplored environment, as well as efficient
victim search performed by humans, requires to incrementally build a map from observations
while keeping track of positions at the same time. This problem is generally
referred to as Simultaneous Localization And Mapping (SLAM) and can be decomposed
into a state estimation problem and a data association problem. The techniques
described in Chapter 3 solve the state estimation problem by continuously tracking the
pose of pedestrians and robots. These methods suffer from the problem of error accumulation,
i.e. the variance of pose estimates increases according to the length of the
traveled trajectory. Therefore, it is required to recognize previously visited places in
order to correct the trajectory estimated by the pose tracker.
However, in unstructured environments, data association is still a challenging problem.
Many existing techniques have been developed under strong assumptions, for
example, they require polygonal structures, as they are typically found in office-like
environments [Gutmann and Schlegel, 1996, Grisetti et al., 2002] or depend on predictable
covariance bounds from pose tracking for solving the data association problem
by validation gating [Dissanayake et al., 2001]. For example, it is not guaranteed
that LRF-based methods work reliably if far-reading measurements or ambiguous patterns
prevent distinguishable features. Data association from camera images requires
at least partially stable illumination conditions, and might lead to unsatisfactory results
in highly dynamic environments. Particularly in the context of disaster response, these
methods can only be applied limitedly. Firefighters at 9/11 reported that they had major
difficulties to orientate themselves after leaving collapsed buildings due to limited
visibility, e.g. caused by smoke and fire, as well as due to the lack of recognizable structures,
such as corridors and doorways. Furthermore, there might be places that are only
accessible by robots, making it necessary to integrate humans and robots into one team
for mapping the area after a disaster. Finally, SLAM works with the principle of map
improvement through loop-closure, however, when facing the reality of emergency response,
firefighters will intentionally try to avoid performing loops, e.g. while they are
searching for victims.
These requirements and extraordinary circumstances make it very hard to apply common
techniques from robotics. An alternative solution to the data association problem
63

64 Chapter 4. RFID Technology-based SLAM
is to utilize RFID technology. RFID tags have a worldwide unique number, and thus
offer a robust way to label and to recognize locations in harsh environments. Their
size is already below 0.5 mm, as shown by the µ-chip from Hitachi [Hitachi, 2003],
and their price is lower than 13¢ [AlienTechnology, 2003]. Passive RFID tags do not
require to be equipped with a battery since they are powered by the reader if they are
within a certain distance. Their reading and writing distance, which depends on the employed
communication frequency, can be assumed to be within a range of meters (see
Chapter 2).
In this chapter, RFID-SLAM, a novel method for SLAM is contributed. The method
utilizes RFID technology for data association, allowing robust and efficient loop closure
in large-scale environments. Specifically the RFID approach enables information
sharing between pedestrians and robots, facilitating it for individual team members
to improve their map without performing loops. Furthermore, a decentralized variant
of RFID-SLAM (DRFID-SLAM) is proposed, which allows agents to jointly improve
their maps without requiring radio communication.
In the proposed approach, RFIDs are actively deployed by robots or humans at adequate
locations, as for example narrow passages that are likely to be passed. The
displacements between RFID tags are estimated by pose tracking methods, and utilized
for building a joint RFID graph from multiple human and robot trajectories, which is
globally optimized by minimizing the Mahalanobis distance [Lu and Milios, 1997].
The emerging graph structure can be considered as a topological map containing metric
information describing the displacement between the nodes. The advantage is that
robots and humans are able to compute their global metric position efficiently based on
a sparse graph representation. Finally, the jointly corrected RFIDs are used as global
constraint points for interpolating the trajectory of each single agent. Pose tracking on
robots is carried out from slippage-sensitive wheel odometry and IMU data (see Chapter
3, Section 3.3), and pose tracking by humans is based on the IMU sensor only (see
Chapter 3, Section 3.5).
The introduced method was evaluated extensively by human and robot experiments,
which were performed indoors and outdoors. The results show that the method is capable
of closing large loops within a few seconds, and moreover, allows to correct
loop-free trajectories if they are shared between the agents.
The remainder of this chapter is organized as follows. In Section 4.2 conventional
SLAM approaches are addressed. Centralized RFID-SLAM is discussed in Section 4.3,
and the distributed version of the approach (DRFID-SLAM) will be presented in Section
4.4. In Section 4.5 results from experiments are proposed, and in Section 4.6 related
approaches are discussed. Finally, in Section 4.7 the conclusion is presented.

4.2. Conventional techniques 65
4.2 Conventional techniques
4.2.1 EKF-based SLAM
Extended Kalman Filter (EKF) based SLAM is a well-known method for continuously
updating the pose of a robot from motion commands, known as the prediction step, and
landmarks observations, known as the observation step [Bailey, 2002, Durrant-Whyte
et al., 1996, Dissanayake et al., 2001]. The prediction step deals with the motion of
the robot by incrementally applying pose tracking (see Chapter 3), which continuously
increases the uncertainty of the pose estimate according to the error model of the odometry.
The observation step occurs if a landmark is detected in the environment. This step
improves the overall state estimate if a previously stored landmark can be re-observed,
whereas when a landmark is observed for the first time, it is added to the memory
through an initialization process called state augmentation.
The basic idea behind EKF-based SLAM is to correlate the pose of the robot with the
locations of landmarks by storing both in a single state vector and covariance matrix. It
has been shown by the work of Durrant-Whyte that particularly the correlations between
landmarks formulated within the single state vector contribute to the convergence of the
method [Durrant-Whyte et al., 1996]. Under the assumption that landmarks are static
in the environment, their correlation improves monotonically as more observations are
made [Dissanayake et al., 2001].
Let the pose of the robot be the vector l = (x, y, θ) T with 3 × 3 covariance matrix Σ l
and the locations of n landmarks by the vector m = (x 1 , y 1 , x 2 , y 2 , · · · , x n , y n ) T with n × n
covariance matrix Σ m , then the single state vector s is defined by:
( ) l
s =
(4.1)
m
( )
Σl Σ
Σ s =
lm
. (4.2)
Σ ml Σ m
It is assumed that at the beginning of the recursive procedure the robot starts at location
(0, 0) with an empty set of observed landmarks, i.e. l = 0 and Σ s = Σ l = 0. At each
motion update the prediction step is carried out by applying pose tracking according to
the input u t = (d t , α t ) with covariance matrix Σ u :
( ) F (lt−1 , u
s t = G (s t−1 , u t ) =
t )
(4.3)
m t−1
Σ st = ∇G s Σ st−1 ∇G T s + ∇G u Σ u ∇G T u , (4.4)

66 Chapter 4. RFID Technology-based SLAM
where
( ) ∇Fl 0
∇G s =
(4.5)
0 I
( ) ∇Fu
∇G u = , (4.6)
0
and F lu , ∇F l , and ∇F u are defined by Equations 3.3, 3.6, and 3.7, respectively. Note
that this update does not affect the estimated locations of landmarks m and covariance
matrix Σ m since landmarks are assumed as stationary. However, the update modifies the
partial covariance matrix Σ l representing the uncertainty of the robots pose, as well as
its cross-correlations Σ ml and Σ lm with landmark locations.
From an observation z = (r, φ) of a landmark within range r and bearing φ with 2 × 2
covariance matrix Σ z , the state vector is updated as follows: first, the observation is
associated to one of the landmarks stored in the state vector. This is carried out by either
utilizing a similarity measure, e.g. based on features detected from the observation,
or by selecting the landmark which is closest to the estimated global location of the
observation, based on the Mahalanobis distance, also known as validation gating. Note
if the landmark is unknown, i.e. cannot be associated to a known one, the state vector
is augmented with the new observation. Second, based on the current estimates of
associated landmark m i = (x i , y i ) and robot pose l = (x, y, θ), the observation is predicted
by the following measurement function:
⎛ √
(x i − x) 2 + (y i − y) 2
H i (s) = ⎜⎝
tan −1 ( y i −y
x i −x
)
− θ
⎞⎟⎠ . (4.7)
Third, the state vector is updated from the observation. This is carried out by computing
the innovation v and its 2 × 2 covariance matrix Σ v by applying the law of error
propagation:
v i = z − H i (s) (4.8)
Σ vi = ∇H s Σ s ∇H T s + Σ z , (4.9)
where the Jacobian ∇H s is given by:
( −
∆x
∇H s =
d
∆y
d 2
− ∆y
∆x
0 0 · · · 0
d d
− ∆x −1 0 · · · 0 − ∆y
d 2 d 2
∆y
)
0 · · · 0
d
∆y
(4.10)
0 · · · 0
d 2
with ∆x = x i − x, ∆y = y i − y, and d = √ ∆x 2 + ∆y 2 . Whether an association is accepted
or rejected is determined by computing the Mahalanobis distance between the measured

4.2. Conventional techniques 67
and predicted observation:
M i = v T i Σ −1
v i
v i , (4.11)
An observation is accepted if M i < λ, where λ denotes the gate threshold. If an observation
is accepted, the following Kalman update is applied:
where the Kalman gain K is given by:
s t = s t−1 + Kv i (4.12)
Σ st = Σ st−1 − KΣ v K T , (4.13)
K = Σ s ∇H s Σ vi . (4.14)
The computation of the covariance matrix Σ s requires O ( n 2) operations, where n denotes
the number of landmarks. Hence, the computational cost increases quadratically
with each new landmark the robot observes. Due to the correlations between landmarks
and the pose of the robot, the map continuously improves by each observation.
However, convergence is only guaranteed in the linear case, and fails if the linearizion
performed by the Jacobian matrices is inaccurate, for example if the robot turns around
quickly. However, the major drawback of the procedure is its sensitivity to incorrect
data associations of landmarks [Neira and Tard, 2001]. Suppose the robot follows a
larger loop in the environment and travels for a long distance without re-observing
landmarks. Consequently, the uncertainty of the pose estimate increases by continuously
applying the prediction step. If re-observing a landmark at the end of the loop,
data association might fail since the observation at the predicted pose lies not within
the Mahalanobis distance of the landmark stored in the state vector. Therefore, if the
robot’s pose is highly uncertain compared to the density of re-observable landmarks,
loop-closure and hence convergence of the map cannot be guaranteed.
For the method as it has been described so far, it is assumed that observations take
place sequentially, and if there are multiple observations at a discrete time step t, updates
from observations are performed after a sequence, one update for each observation. Unfortunately,
this strategy does not reflect landmark correlations for specific locations.
Bailey introduced a batch update procedure that significantly reduces the data association
problem by correlating landmarks observed at a single pose [Bailey, 2002]. He has
shown by an empirical evaluation that batch updates lead to more reliable data association
even under extremely noisy pose estimates from the robots odometry. However,
landmarks are associated based on features detected by a laser range finder requiring a
minimal amount of structure found in the environment.

68 Chapter 4. RFID Technology-based SLAM
4.2.2 FastSLAM
A remarkable reduction of the computational complexity of EKF-based SLAM is achieved
by the FastSLAM method, which was introduced by Montemerlo and colleagues [Montemerlo
et al., 2002]. Their approach is based on the work of Murphy [Murphy, 2000],
who observed that, if the trajectory of the robot is known, the problem of determining
the landmark locations can be decomposed into independent estimation problems, one
for each landmark. Moreover, this factored representation is exact since landmark observations
taken by the robot are conditionally independent. Consequently, FastSLAM
solves the SLAM problem by solving collections of independent landmark estimation
problems conditioned on individual trajectory estimates of the robot. The individual
trajectory estimates are maintained by a particle filter, where each particle represents
one possible map of the environment, consisting of a trajectory and N Kalman filters
for estimating the locations of N landmarks. Hence, the computational complexity is
O (MN), where M is the number of particles and N the number of landmarks. They also
introduced a more efficient implementation based on a tree-like data structure, leading
to a computational complexity of O ( M log K ) .
Another advantage of the method is its ability to track multiple hypotheses of data
associations. In contrast to EKF-based SLAM, which decides once the association between
a predicted and known landmark by the validation gate shown by Equation 4.11,
FastSLAM allows to track simultaneously multiple hypotheses, each represented by a
particle, whereas unlikely associations are filtered out by a re-sampling process. However,
this method also might fail to converge given an arbitrary noisy pose estimate of
the robot. Furthermore, it is unclear how many particles have to be utilized under harsh
environmental conditions as they are present in areas after a disaster. In the next section
it will be shown that RFID technology offers a straight-forward solution to the data
association problem.
4.3 Centralized RFID-SLAM
RFID-SLAM [Kleiner et al., 2006c] is a procedure for mapping an area based on the
odometry trajectories and RFID observations from multiple robots and humans in the
field. For the sake of simplicity, robots and humans are denoted as “field agents”,
which communicate their observations to a central station. More specifically, the agents
estimate distances between RFID locations by the pose tracking methods described
in Chapter 3, and communicate these estimates back to the station, which centrally
combines and corrects all trajectories. Furthermore, it is assumed that RFID detections
are within a range below one meter allowing to cover corridors and doorways while
providing sufficient positioning accuracy.
From the correspondences of observed RFID tags and their estimated displacements,

4.3. Centralized RFID-SLAM 69
Figure 4.1: The spring-mass analogy to the generated RFID graph. Vertices represent
RFIDs (masses) and edges between them represent measured distances with covariances
(springs).
the station computes the globally consistent RFID locations by minimizing the Mahalanobis
distance [Lu and Milios, 1997]. These corrected locations are finally used as
global constraint points for interpolating single agent trajectories. The global optimization
method can be illustrated by considering the analogy to a spring-mass system (see
Figure 4.1). Consider the locations of RFIDs as masses and the measured distances
between them as springs, where the uncertainty of measurements corresponds to the
elasticity of the springs. Then, finding a globally consistent map is equivalent to finding
an arrangement of the masses that requires minimal energy.
The reminder of this section is structured as follows. In Section 4.3.1 the construction
of RFID graphs from communicated displacement estimates is discussed. The linear
and non-linear optimization of RFID graphs is shown in Section 4.3.2 and Section 4.3.3,
respectively. In Section 4.3.4 the interpolation of trajectories, and in Section 4.3.5 the
RFID sensor model will be introduced.
4.3.1 Building RFID graphs
Each time an RFID has been observed, a message is generated that contains the ID of
the previously visited RFID i and the currently visited RFID j, as well as an estimate
of the local displacement between both RFIDs. The local displacement is estimated by
a Kalman filter, integrating data from pose tracking (Chapter 3). If an RFID has been
detected, the Kalman Filter is reset in order to estimate the relative displacement to the
subsequent tag on the trajectory. The noisy measurement of the local displacements is
denoted by ˆd i j = d i j + ∆d i j . It is assumed that the error ∆d i j is normally distributed and

70 Chapter 4. RFID Technology-based SLAM
thus can be modeled by a Gaussian distribution with zero mean and 3 × 3 covariance
matrix Σ i j . The local displacement ˆd i j is defined by the vector (∆ ˆx i j , ∆ŷ i j , ∆ˆθ i j ) T , where
∆ ˆx i j and ∆ŷ i j denote the relative spacial displacement, and ˆθ i j the relative orientation
change.
The central station incrementally builds a global graph from all displacement estimates
communicated by the field agents through utilizing the unique ID of RFIDs for
data association. The constructed graph G = (V, E) consists of vertices V and edges E,
where each vertex represents an RFID tag, and each edge ( V i , V j
)
∈ E represents an estimate
ˆd i j with covariance matrix Σ i j between two RFID tags associated with vertices V i
and V j , respectively. The graphs from all field agents are unified in the following way:
on the one hand, if the same vertex has been observed twice, a loop has been detected
on the graph. A detected loop is modeled by a pseudo edge between the same RFID
node with respect to the RFID antenna’s sensor model, which will be further described
in Section 4.3.5. On the other hand, if two or more field agents observe the same edge,
i.e. their trajectory overlaps between two or more neighboring RFIDs, both observations
are merged by an Extended Kalman Filter (EKF) [Maybeck, 1979].
We denote the true pose vectors of n + 1 RFID nodes by l 0 , l 1 , . . . , l n , and the function
calculating the true displacement (∆x i j , ∆y i j , ∆θ i j ) T between a pair of nodes ( )
l i , l j
is denoted as measurement function d i j . The goal is to find the true locations of the l i j
given the set of measurements ˆd i j and covariance matrices Σ i j . This can be achieved according
to the maximum likelihood concept by minimizing the following Mahalanobis
distance:
l 0:n = arg min
l 0:n
∑
i, j
( ) T ( )
di j − ˆd i j Σ
−1
i j di j − ˆd i j , (4.15)
where l 0:n denotes the concatenation of poses l 0 , l 1 , . . . , l n . Since measurements are taken
relatively, it is assumed without loss of generality that l 0 = 0 and l 1 , · · · , l n are relative
to l 0 . Moreover, the graph is considered as fully connected, and if there does not exist a
measurement between two nodes, the inverse covariance matrix Σ −1
i j
is set to zero.
4.3.2 Linear optimization
If poses are modeled without orientation angle θ, then the optimization problem can
be solved linearly. Assume the two connected RFID locations l i = (x i , y i ) T and l j =
(x j , y j ) T , where l i follows l j on the trajectory. Then, Equation 4.15 can be solved by
inserting d i j = l i − l j :
l 0:n = arg min
l 0:n
∑
i, j
( ) T ( )
li − l j − ˆd i j Σ
−1
i j li − l j − ˆd i j . (4.16)
In order to solve the optimization problem analytically, Equation 4.16 has to be transformed
into matrix form. The measurement functions for all edges can be described by

4.3. Centralized RFID-SLAM 71
the following matrix equation:
d = hl, (4.17)
where vector l denotes the concatenation of pose vectors l 0 , . . . , l n with the dimension
nd, vector d denotes the concatenation of all pose differences d i j = l i − l j . and matrix
h is an incidence matrix with elements {−1, 0, 1}. For example, a fully connected graph
consisting of the three nodes l 0 , l 1 , l 2 has the following form:
⎛
d 01
d 02
d 10
d 12
d 20
⎜⎝
⎞⎟⎠
d 21
⎛ ⎞
−1 0
0 −1
1 0
=
1 −1
0 1 ⎜⎝ ⎟⎠
−1 1
Consequently, Equation 4.16 can be rewritten in matrix form by
(
l1
l 2
)
. (4.18)
l = arg min
l
(ˆd − hl ) T
Σ
−1
(ˆd − hl ) , (4.19)
ˆd
where hl denotes the measurement function as described by Equation 4.17, ˆd denotes
the concatenation of observations ˆd i j , and Σ −1 denotes the inverse covariance matrix of
ˆd
ˆd, consisting of the inverse sub-matrices Σ i j . Finally, the minimization problem can be
solved by
and the covariance of l can be calculated by
l = ( h T Σ −1
ˆd h) −1
h T −1
Σˆd
ˆd , (4.20)
Σ l = ( h T Σ −1
ˆd h) −1
. (4.21)
As shown by Lu and Milios, the computation of Equations 4.20 and 4.21 can be simplified
under the assumption of statistical independence between distance measurements
d i j . Consequently, the nd × nd matrix h T Σ −1 h can be replaced with the matrix g, which
ˆd
can be computed by:
g ii =
n∑
j=0
Σ −1
i j , i = 1, . . . , n,
g i j = −Σ −1
i j , i = 1, . . . , n , j = 1, . . . , n , i j.
(4.22)
Note that Σ i j = Σ ji , and the d ×d matrices Σ i j have to be invertible. Furthermore, the nddimensional
vector h T Σ −1
l
ˆd can be replaced by the vector b, where the d-dimensional

72 Chapter 4. RFID Technology-based SLAM
sub-vectors can be computed by
b i =
n∑
j=0; ji
Σ −1
i j
ˆd i j , i = 1, . . . , n. (4.23)
Note that ˆd i j = ˆd ji . Finally, the position estimate and covariance can be written as
l = g −1 b (4.24)
Σ l = g −1 (4.25)
Equation 4.24 and Equation 4.25 can be solved in O ( n 3) , where n is the number of
nodes, since the nd × nd matrix g has to be inverted. However, in practice, this computation
can be performed much more efficiently since matrix g is a sparse matrix. This
is due to the fact that only a little amount of nodes in the graph are connected with
each other ∗ . Therefore, most covariances Σ i j are set to zero. There are numerous packages
available that have been developed for the efficient computation of sparse matrices,
as for example the Sparse matrix multiplication package (SMMP) [Bank and Douglas,
1993].
4.3.3 Non-linear optimization
The method described in Section 4.3.2 requires that the measurement functions d i j , and
the measurements dˆ
i j are linear. However, particularly in the context of robot localization,
poses are typically modeled with an angular component, e.g. l i = (x i , y i , θ i ) T . One
“trick” in Kalman filtering based methods is to linearize the measurement function with
a Taylor series. After the linearization, Equations 4.24 and 4.25 from Section 4.3.2 can
be applied in the same way as if the measurement function would be linear † . Therefore,
the goal is to determine the linearized measurement functions d
i ′ j
and linearized
d ′ i j
measurements d ˆ
i ′ j
, in order to solve the optimization problem by Equation 4.24. Note
that the measurements ˆ are already linearized if they have been computed by EKFbased
pose tracking, as described in Chapter 3. In the following, the linearization of the
measurement function will be discussed.
Assume the two connected locations l i = (x i , y i , θ i ) T and l j = (x j , y j , θ j ) T with the
local displacement d i j = (∆x, ∆y, ∆θ) T , where l i follows l j on the trajectory. Then, the
displacement between these locations can be determined by the following transforma-
∗ In the context of robot navigation node connections are naturally constrained by the environment.
† Under significant angular changes the linearization procedure might lead to errors, requiring the optimization
process to be applied iteratively.

4.3. Centralized RFID-SLAM 73
tion [Smith and Cheeseman, 1986]:
which will be denoted by
∆x = (x i − x j ) cos θ j + (y i − y j ) sin θ j (4.26)
∆y = −(x i − x j ) sin θ j + (y i − y j ) cos θ j (4.27)
∆θ = θ i − θ j , (4.28)
d i j = l i ⊖ l j . (4.29)
Let ∆d = ˆd i j −d i j denote the observation error, which is a function of the two poses l i and
l j since d i j = l i ⊖ l j . It is assumed that the estimates ˆl i = ( ˆx i , ŷ i , ˆθ i
)
and ˆl j = ( ˆx j , ŷ j , ˆθ j
)
of
these poses are given from pose tracking. Then, the observation error can be linearized
by the following Taylor series [Lu and Milios, 1997]:
∆d i j ≈ ˆd i j − (ˆl i ⊖ ˆl j ) + ˆK −1
j (∆l i − Ĥ i j ∆l j ), (4.30)
where
⎛
⎞
cos ˆθ j sin ˆθ j 0
ˆK −1
j = − sin ˆθ ⎜⎝ j cos ˆθ j 0⎟⎠ 0 0 1
(4.31)
⎛
⎞
1 0 −ŷ i + ŷ j
Ĥ i j = 0 1 ˆx ⎜⎝
i − ˆx j ⎟⎠ , 0 0 1
(4.32)
∆l i = ˆl i − l i ; ∆l j = ˆl j − l j (4.33)
If replacing matrix Ĥ i j with the product Ĥi
−1 Ĥ j , Equation 4.30 can be transformed to
−Ĥ i ˆK j ∆d i j = Ĥ i ˆK j ((ˆl i ⊖ ˆl j ) − ˆd i j ) − (Ĥ i ∆l i − Ĥ j ∆l j ), (4.34)
which is equal to
if substituting the following terms
∆d ′ i j = ˆd ′ i j − d ′ i j, (4.35)
∆d i ′ j = −Ĥ i ˆK j ∆d i j (4.36)
ˆd i ′ j = Ĥ i ˆK j ((ˆl i ⊖ ˆl j ) − ˆd i j ) (4.37)
d ′ i j = Ĥ i ∆l i − Ĥ j ∆l j . (4.38)

74 Chapter 4. RFID Technology-based SLAM
Hence, the linearized measurement function d
i ′ j with covariance matrix Σ′ i j
is given by:
d ′ i j = Ĥ i ∆l i − Ĥ j ∆l j (4.39)
Σ ′ i j = Ĥ i ˆK i Σ i j ˆK T j Ĥ T i (4.40)
(4.41)
which can be utilized for formulating the linearized optimization problem
∑ ( )
l = arg min d
′
i j − ˆd i ′ T ( )
j Σ
′−1
i j d
′
i j − ˆd i ′ j , (4.42)
l
i, j
and can be solved in the same way as proposed in Section 4.3.2. Since the linearization
leads to errors, the procedure has to be applied iteratively, where the result from the
nth computation serves as a starting point for the n + 1th computation. After the optimization
has been performed, the coordinates of the optimal poses can be determined
as follows: assume l
i ′ and Σ′ i
are the poses and covariances resulting from the linearized
optimization. Then, the final locations of the corrected poses can be determined by:
l i = ˆl i − Ĥ −1
i l ′ i (4.43)
Σ i = (Ĥi
−1 )Σ ′ i(Ĥi −1 ) T (4.44)
4.3.4 Trajectory interpolation
The corrected RFID network can be used as a basis for correcting the odometry trajectory
of each agent. This is carried out by utilizing the corrected RFID locations
as constraint points for the correction of the trajectory. One possibility to do this, is
to augment the noisy odometry trajectories with these constraints, e.g. by connecting
them with pseudo edges, and to correct them once more with the method that has been
described in the last section. However, since these constraint points are already globally
consistent, it turns out to be much more efficient to perform a local interpolation of
trajectory poses between the corrected RFIDs.
Given a sequence of corrected RFID locations r 1:t and an uncorrected trajectory x 1:t
the corrected trajectory y 1:t is computed by interpolating each pose x k ∈ x 1:t between the
two closest RFIDs. In order to correct each pose x k of the trajectory, we first determine
the corrected RFID locations r i and r j of the two closest RFIDs before and after the pose,
with j ≥ k ≥ i, and the uncorrected poses x i and x j at corresponding time, respectively.
Finally, the corrected pose is computed by:
( ( ))
w1 (x i − r i ) + w 2 x j − r j
y k = x k −
w 1 + w 2
, (4.45)

4.4. Decentralized RFID-SLAM (DRFID-SLAM) 75
where weights w 1 and w 2 are computed by w 1 = ∣ ∣ ∣r j − x k
∣ ∣∣ and w2 = |r i − x k |. Experimental
results in Section 4.5 will show that this method together with the RFID graph
optimization allows to efficiently and robustly correct large trajectories.
4.3.5 Sensor model
Each time a loop has been detected on the trajectory, i.e. an RFID has been observed
twice, a pseudo edge is added to the corresponding RFID node. We model this edge by
accounting for the spatial expansion of the utilized RFID antenna.
For experiments, two different RFID antennas, one for humans and one for robots,
were utilized. The antenna of the robot was mounted in parallel to the ground allowing
to detect RFIDs lying beneath the robot, whereas the antenna of the human was carried
manually. An evaluation of the robot RFID system has shown that RFID tags lying
beneath the robot within the antenna’s expansion (a rectangle of ≈ 20 × 15 cm), are
constantly detected, also if the robot travels with high velocity. The antenna carried
by the human detects RFIDs within a range of approximately 30 cm if the antenna is
aligned to the bearing of the RFID tag. However, it is not possible to tell the exact
position of the detection within this range. Consequently, the observation uncertainty is
modeled according to the maximal detection range.
It is assumed that, in the average case, RFIDs are detected within the antenna’s center,
and that RFID detections can occur at arbitrary orientations of the human or robot.
Therefore, a “loop-closing” edge is modeled with distance ˆd ii set to (0, 0, ∆θ) T , where
∆θ denotes the angle difference between the two pose estimates of the RFID. Under
the assumption that RFIDs are detected at 95% probability if they are within maximal
reading range d M , we model the according covariance matrix by:
⎛
⎞
2 2 d 2 M
0 0
Σ ii = 0 2 ⎜⎝
2 d 2 M
0 ⎟⎠ , (4.46)
0 0 σ 2 θ
where σ 2 θ is the linearized variance of the angle, and 22 d 2 M
the variance of the normal
distribution over the interval [−2d M ; 2d M ]. Note that σ 2 θ
has to be linearized according
to the procedures described in Section 4.3.3. Finally, The robot antenna was modeled
with d M = 20 cm, and the antenna carried by the human with d M = 30 cm.
4.4 Decentralized RFID-SLAM (DRFID-SLAM)
In emergency response scenarios, communication links between agents or to a central
station might be cut-off temporarily or constantly. One obvious reason is that in building
structures radio waves are not able to penetrate walls if they are made of reinforced

76 Chapter 4. RFID Technology-based SLAM
concrete. Otherwise, it might also be likely that first responders, which search for
victims on large terrain, are simply out of communication range. Moreover, they might
visit the same location at different time, and thus are also not able to communicate.
Therefore, methods deployed in this context have to be robust towards total loss of
communication. Decentralized RIFD-SLAM (DRFID-SLAM) is particularly designed
for situations in which radio communication is impossible. Note that this distinguishes
the approach from other methods that require at least low-bandwidth communication
channels between the agents [Reece and Roberts, 2005,Nettleton et al., 2003,Kim et al.,
2004]. Furthermore, we assume that RFIDs have sufficient memory for storing RFID
graph structures as described in Section 4.3.
The basic idea of DRFID-SLAM is to utilize the memory of RFID tags for learning
the topology of the surrounding RFID graph. When agents pass a tag, they update
their knowledge on the graph topology by reading data from the tag’s memory, as well
as synchronize the tag’s memory by writing graph data they have collected on their
trajectory. By this, agents are able to learn a graph larger than their own trajectory, i.e.
extended by trajectories of other agents. The union of single agent trajectories leads
to graphs that contain loops, hence allowing agents to correct their trajectories, even if
they do not contain loops. Figure 4.2 depicts an example of DSLAM of three agents
successively passing RFIDs. The third agent is finally able to close a loop consisting of
partial tracks from the first and second agent.
R 3
R 1
R 2
R 3
R 1
R 2
R 3
R 1
R 2
R 4
R 5
R 6
R 7
R 9
R 4
R 5
R 6
R 7
R 9
R 4
R 5
R 6
R 7
R 9
R 8
R 8
R 8
First Agent
Second Agent
Third Agent
Figure 4.2: Example for decentralized SLAM of three agents: RFID nodes
R 1 , R 2 , · · · , R 9 successively learn the trajectories of the passing agents. The third agent
learns at node R 1 the trajectories from the first and second agent, enabling loop-closure
for map improvement when reaching node R 6 .
The procedure can be considered as a step-wise information propagation through the
network of RFIDs, where each visited RFID learns about nearby RFID nodes from
the trajectory of its visitor, as well as from trajectories that were previously stored by
other agents in the nodes passed by the visitor. Therefore, after a sufficient amount of
trajectories has been traveled, a single RFID node memorizes the whole infrastructure

4.4. Decentralized RFID-SLAM (DRFID-SLAM) 77
of the network. Note that for supporting map improvements, RFIDs do not necessarily
have to store the whole network structure.
One problem arising in the context of distributed SLAM is the double counting of
information propagated via different paths, i.e. rumor propagation. While this problem
typically requires to apply probabilistic methods, such as Bounded Covariance Inflation
(BCI) [Reece and Roberts, 2005], and Covariance Intersection (CI) [Nettleton et al.,
2003], it can be solved rather trivially with RFIDs. Here the basic idea is to manage
displacement estimates locally in the memory of RFIDs, i.e. to distinguish each estimate
by location and time where it has been observed.
We distinguish edges as managed and propagated in the memory of RFIDs and introduce
a counter C i attached to each RFID, which is locally updated each time an
agent writes information concerning a managed edge. Edges that are managed by node
i are all the adjacent edges that have i in common, i.e. those edges that directly connect
to i. Consider the case that an agent travels from RFID i to RFID j and generates
thereby a new estimate ( ˆd i j , Σ i j
)
. Then, RFID j is the managing node, i.e. manages
the edge between i and j. Consequently, the agent stores in RFID j a new entry
e i j = 〈 i, j, ˆd i j , Σ i j , C i
〉
consisting of the IDs of the two visited RFIDs i and j, their displacement
estimate, and the current count value of the RFID, which is incremented after
writing. Obviously, there can be two nodes managing the same edge since agents can
either travel from i to j or from j to i. Therefore, we have to distinguish both cases
during propagation, which is solved by defining 〈i, j, · · · 〉 〈 j, i, · · · 〉, i.e. the managing
node corresponds to the second node in the sequence.
Algorithm 2: Updating the graph R k of visited RFID k
Input: New entry e ik , a set of graphs S = { R j : j ∈ P } , where P denotes the RFID trajectory of
the agent
Output: Updated graph R k
Update managed edges:
R k ← R k ∪ 〈 i, k, ˆd ik , Σ ik , C k
〉
;
C k ← C k + 1;
Update propagated edges:
foreach R j ∈ S do
foreach e ∈ R j do
if e ∈ R k then
continue;
R k ← R k ∪ e;
end
end
Propagated edges are managed edges after they have been copied by agents. If visiting
a node, the agent performs a union operation between all entries stored in the memory
of the RFID, denoted by graph R i , and its own graph, denoted by A i , and propagates

78 Chapter 4. RFID Technology-based SLAM
them further to other RFIDs. The double counting of identical estimates is prevented
by ensuring that there always exists only one unique entry for each combination of i, j,
and C j and unique sequence < i, j > within each RFID, where C j is the counter value of
RFID j during the creation of the managed edge. The complete procedure for updating
RFID nodes is shown by Algorithm 2. Note that the agents update their own graphs A i
by the same procedure, while considering propagated edges only.
The decentralized information exchange facilitates the building of RFID graph structures
as described in Section 4.3. Hence, agents are able to improve their trajectories if
they detect a new loop after performing the union procedure. However, the possibility
to close loops from joining trajectories depends on the number of agents that visit a
node, as well as on the sequence trajectories have been traveled. For example, in the
situation depicted in Figure 4.2, the first agent does not benefit from the tracks of the
others. The performance of the proposed approach can be further improved by extending
the information propagation via radio communication. Each time agents are within
communication range, they can unify their graphs, i.e. to exchange propagated edges,
by the procedure described in Algorithm 2. Thereby, they are possibly able to connect
large graphs, accelerating the map joining process.
4.5 Experimental results
Experiments on centralized and decentralized RFID-SLAM were conducted with a team
of humans, a team of robots, and finally jointly by humans and robots. All experiments
were carried out in both indoor and outdoor scenarios, where the Zerg robot platform,
described in Chapter 2, has been deployed.
4.5.1 Experimental setup
The robot platform utilized for experiments is equipped with an RFID system consisting
of an RFID reader and antenna. The active distribution of RFID tags is carried out by a
custom-built actuator based on a metal slider that can be moved by a conventional servo
(Figure 4.3(b)).
The slider is connected with a magazine that maximally holds around 50 tags. Each
time the mechanism is triggered, the slider moves back and forth while dropping a
single tag from the magazine. The device is constructed in a way that for each trigger
signal only one tag is released. A software module triggers the device at adequate
locations, which are determined according to the existing density of RFIDs, i.e. maintaining
a maximal defined density of RFIDs, and also, if operating in indoor scenarios,
with respect to the structure of the environment. For example, narrow passages, such
as doorways and corridors, are likely to be passed by the robot, thus RFIDs are deployed
with high probability in these kind of environmental structures. The width of

4.5. Experimental results 79
(a) (b) (c)
Figure 4.3: A novel mechanism for the active distribution of RFID tags. (a) The utilized
RFID tags. (b) The mechanism with servo. (c) The mechanism together with the
antenna mounted on the Zerg robot.
the free space surrounding the robot is computed from the distance between the obstacles
located most left and most right to the robot. They are found on a line which goes
through the center of the robot, and which is orthogonal to the robot’s orientation.
The antenna of the reader is mounted parallel to the ground, allowing to detect RFIDs
lying beneath the robot. In order to enable the robot to perceive the deployment of
RFIDs, the deploy device has been mounted above the RFID antenna, forcing deployed
RFIDs to pass directly through the antenna. We utilized Ario RFID chips from Tagsys
(Figure 4.3 (a)) with a size of 1.4 × 1.4 cm, 2048 Bit RAM, and a response frequency
of 13.56 MHz. For the reading and writing of these tags, we employed a Medio S002
reader, likewise from Tagsys, which operates within a range of approximately 30 cm
while consuming less than 200 mA. Figure 4.3 (c) shows the complete construction
consisting of deploy device and antenna mounted on the Zerg robot. For a detailed
description of the RFID sensor model see Section 4.3.5.
Pose tracking on the Zerg platform was carried out by slippage-sensitive odometry,
as described in Section 3.3, and the Xsens MTi IMU, as described in Chapter 2. For
pose tracking on pedestrians, we utilized the MTx IMU from Xsens, which has been
attached to a test person for measuring the vertical and lateral acceleration, as well as
the azimuth orientation (see the orange box in Figure 3.7 (a)). Based on an empirical
evaluation, we modeled pose tracking uncertainty on pedestrians with σ 2ˆd = (0.05 m)2 d,
and σ 2ˆθ = (15◦ ) 2 , and on robots with σ d = 0.816 cm m and σ d slip
= 24.72 cm m .
During outdoor experiments, ground truth data was obtained by a GNSS device. For
this purpose, we used the GPSlim236 GPS receiver from Holux, which is equipped with
SIRFstar III technology. The receiver allows to track up to 20 satellites at an update
rate of 1 Hz and has a position accuracy of 5 − 25 meters without DGPS. Furthermore,
the receiver is able to process data from the EGNOS system ‡ , improving the horizontal
‡ EGNOS stands for European Geostationary Navigation Overlay Service and provides Differential GPS
(DGPS) via satellites that broadcast correction signals on the same frequency as GNSS satellites.

80 Chapter 4. RFID Technology-based SLAM
position accuracy towards < 2.2 meters and vertical position accuracy towards < 5
meters at 95% of the time. Note that we were able to also obtain ground truth in semiindoor
scenarios since RFIDs where intentionally placed outdoors.
4.5.2 RFID-SLAM with robots
The described method for RFID technology-based SLAM was tested extensively with
data generated by a simulator [Kleiner and Buchheim, 2003] and on the Zerg robot
platform. The simulated robot explored three different building maps, the small map,
normal map, and large map of the sizes 263 m 2 , 589 m 2 1240 m 2 , while automatically
distributing RFID tags in the environment. Figure 4.4 (a-c) shows the averaged results
from 100 executions of RFID-SLAM on the three maps at five different levels of odometry
noise. We measured a computation time of 0.42 seconds on the small map, 2.19
Accuracy [mm]
4000
3500
3000
2500
2000
1500
1000
500
0
1
Odometry
RFID corrected
2
Small map
3
Noise level (X times of odometry noise)
(a)
4
5
Accuracy [mm]
4000
3500
3000
2500
2000
1500
1000
500
0
1
Odometry
RFID corrected
2
Normal map
3
Noise level (X times of odometry noise)
(b)
4
5
Accuracy [mm]
4000
3500
3000
2500
2000
1500
1000
500
0
1
Odometry
RFID corrected
2
Large map
3
Noise level (X times of odometry noise)
(c)
4
5
Figure 4.4: (a) - (c) Result from applying RFID-based SLAM at different levels of
odometry noise on (a) the small map (263 m 2 ), (b) the normal map (589 m 2 ), and (c) the
large map (1240 m 2 ).
seconds for the normal map, and 13.87 seconds for the large map, with a Pentium4

4.5. Experimental results 81
(a)
(b)
(c)
(d)
(e)
(f)
Figure 4.5: (a,c,e) Result from RFID-SLAM of a robot driving in a cellar: (a) the noisy
map, (c) the corrected map, and (e) the ground truth created by iterative scan matching.
(b,d,f) Result from applying the RFID-SLAM to data generated in the simulation. (b)
The small map with odometry noise, (d) the corrected map, and (f) the ground truth
taken directly from the simulator’s map editor. Note that the cellar’s ground truth map
displays unoccupied rooms. The rectangular structures that can be seen in the upper left
room in the constructed maps (a) and (c) originate from crates stored in those rooms.
2.4 GHz. The small map before and after the correction is shown in Figure 4.5 (b,d). To
achieve this result, the robot distributed approximately 50 RFIDs.
In order to evaluate the performance of RFID-SLAM in a real environment, we collected
data from a robot autonomously exploring a cellar for 20 minutes while detecting
RFID tags on the ground. The robot continuously tracked its pose as described in Sec-

82 Chapter 4. RFID Technology-based SLAM
tion 3.3. As depicted in Figure 4.5, the non-linear method successfully corrected the
angular error based on RFID data association. The correction was based on approximately
20 RFID tags.
Figure 4.6: Result from applying RFID-SLAM outdoors while driving with 1m/s on
a parking lot. Trajectories are visualized with GoogleEarth, showing RFID locations
estimated by the odometry (red), the ground truth (blue), and corrected by RFID-SLAM
(green).
Additionally, we conducted an outdoor experiment with the Zerg robot driving with
an average speed of 1 m/s on a parking lot. The odometry was generated from the wheel
encoders (translation) and IMU (rotation). Furthermore, the robot detected RFIDs with
the antenna described above. We obtained position ground truth from both Differential
GPS (DGPS) and manual measurements, whereas faulty GPS positions, e.g. due
to multi-path propagations close to buildings, were corrected from the manual measurements.
Figure 4.6 shows the RFID locations estimated by the odometry (red), the
ground truth (blue), and corrected by RFID-SLAM (green). The corrected trajectory has
a mean Cartesian error of 1.8 ± 3.1 m, compared to the uncorrected trajectory, which
has a mean Cartesian error of 8.3 ± 8.5 m.
Furthermore, we evaluated the influence of the number of detected RFIDs with respect
to the stability of the SLAM approach. Figures 4.7 (a), (b), and (c) show the
Cartesian error, the cross-track error (XTE), and the along-track error (ATE) at decreasing
number of RFIDs, respectively. The route graph optimization consistently improves
the accuracy of the trajectory, even with a comparably little number of RFIDs, e.g.

4.5. Experimental results 83
Error [m]
25
20
15
10
5
0
18
Corrected
Uncorrected
Cart (Cartesian-Track Error)
9
6
4
Number of RFIDs
(a)
3
2
Error [m]
18
16
14
12
10
8
6
4
2
0
18
Corrected
Uncorrected
9
XTE (Cross-Track Error)
6
4
Number of RFIDs
(b)
3
2
Error [m]
18
16
14
12
10
8
6
4
2
0
18
Corrected
Uncorrected
9
ATE (Along-Track Error)
6
4
Number of RFIDs
(c)
3
2
Figure 4.7: Result from applying RFID-SLAM outdoors while driving with 1 m/s for
approximately 1 km on a parking lot with the Zerg robot. (a) the Cartesian error, (b) the
cross-track error (XTE), and (c) the along-track error (ATE) with respect to the number
of utilized RFIDs.
one RFID each 500m. The correction of 18 RFIDs took 2.1 seconds on a PentiumM
1.7 MHz, and the interpolation of the odometry trajectory took 0.2 seconds.
In order to evaluate the scalability of the approach in large-scale environments, a second
outdoor experiment was conducted. During this experiment, the robot was driving
a total distance of more than 2.5 km with an average speed of 1.58 m/s. Note that this
velocity requires human beings to walk comparably fast in order to follow the robot.
Furthermore, the robot was heavily shaking from fast navigation over uneven ground,
such as road holes, small debris, and grass. Also during this experiment, pose tracking
was carried out based on data from the wheel encoders (translation) and IMU (rotation),
and position ground truth has been obtained from DGPS. The optimization yielded an
average Cartesian error of 9.3 ± 4.9 m, compared to the uncorrected trajectory, which
has an average Cartesian error of 147.1 ± 18.42 m. The correction of 10 RFIDs took 0.3
seconds on a PentiumM 1.7 MHz, and the interpolation of the odometry trajectory took

84 Chapter 4. RFID Technology-based SLAM
Improved odometry trajectory with covariance bounds
Globally corrected trajectory with covariance bounds
450
400
250
350
300
200
Y [m]
250
200
Y [m]
150
150
100
100
50
0
-50
-350
-300
3σ Bound
Odometry
-250 -200 -150 -100 -50
X [m]
(a)
0
50
100
150
200
50
0
-200
3σ Bound
Corrected track
-150
-100 -50
X [m]
(b)
0
50
Figure 4.8: Covariance bounds from applying RFID-SLAM outdoors: (a) before and
(b) after the correction.
Figure 4.9: Result from applying RFID-SLAM outdoors while driving with 1.58 m/s
for more than 2.5 km with a Zerg robot showing the odometry trajectory (red), the
ground truth trajectory (blue), and the corrected RFID trajectory (green).

4.5. Experimental results 85
2.4 seconds. Figure 4.8 shows the covariance bounds during EKF-based dead reckoning
of the improved odometry (a), and after the global optimization (b). Figure 4.9 shows
the trajectory of the odometry (red), the ground truth (blue), and the corrected RFID
graph (green). Note that for the sake of readability, Figure 4.8 only shows the first loop
of the performed trajectory.
4.5.3 RFID-SLAM with humans
In this section experiments conducted by a team of humans and a human-robot team are
shown. Note that multiple trajectories were performed by a single person and data was
recorded for later processing.
Semi-Indoor Experiment
The semi-indoor experiment was carried out on the campus of the University of Freiburg,
including many accessible buildings which were entered by the test person. We measured
that some of these buildings contain magnetic fields disturbing the angle estimate
of the PDR method, as for example, metal stairs or metal doors, such as shown in
Figure 3.7. Figure 4.10 provides an overview of the area, which was generated by
GoogleEarth. During this experiment, the test person traveled six trajectories with
different starting and ending locations, while performing pose tracking with the PDR
method described previously, and while distributing and re-observing RFIDs. In order
to visualize the PDR trajectories, we utilized the accurate starting locations taken from
the ground truth data and projected each PDR trajectory with respect to its starting location
on the map. In Figure 4.10 each trajectory is shown with a different color. Position
accuracy decreases with increasing length of the traveled trajectory.
All trajectories were collected and merged into a single graph for applying the centralized
method described in Section 4.3. The corrected edges between RFIDs are
shown by the black lines in Figure 4.10, as well as the corrected locations of the RFIDs
(small squares). Furthermore, we computed the average Cartesian error with respect to
ground truth. Figure 4.11 depicts these errors according to each pedestrian. It shows
the uncorrected trajectory, the single trajectory corrected on its own, and the trajectory
corrected from the unified graph. The correction based on the unified graph yields best
results for each pedestrian.
Outdoor Experiment
The outdoor experiment was carried out in a larger urban area, containing mainly residential
buildings with up to six floors. Due to the multipath propagation problem in this
environment, e.g. narrow streets and high buildings, the recorded GPS tracks had to be
manually corrected for providing ground truth data. Figure 4.12 (a) depicts the recorded

86 Chapter 4. RFID Technology-based SLAM
Figure 4.10: Result from the semi-indoor experiment: each line indicates the pose
tracking of the trajectory of a single pedestrian. Black lines and squares show the corrected
graph of RFIDs.
50
45
40
Cart (Cartesian-Track Error)
Single corrected
Multi corrected
Uncorrected
35
Error [m]
30
25
20
15
10
5
1
2
3
4
Pedestrian ID
5
6
Figure 4.11: Result from the semi-indoor experiment: the average Cartesian error.

4.5. Experimental results 87
PDR trajectories. During this experiment, the pedestrians walked approximately 9000
meters, while deploying 20 RFID tags. Figure 4.12 (b) depicts the uncorrected RFID
graph constructed by centrally merging all PDR trajectories and RFID observations,
and Figure 4.12 (c) shows the graph after the correction. It can be seen in Figure 4.12
(d) that the corrected graph is close to ground truth. We also computed the average
Cartesian error with respect to ground truth, shown in Figure 4.13. Also during this experiment
the correction based on the unified graph yields better results for each single
pedestrian.
(a)
(b)
(c)
(d)
Figure 4.12: Result from correcting pedestrian trajectories in an urban environment (the
City of Freiburg). (a) Pedestrian trajectories, where each color indicates the recorded
track of a single pedestrian. (b) The uncorrected RFID graph, generated from the PDR
tracks. (c) The corrected RFID graph. (d) Manually improved GPS ground truth.

88 Chapter 4. RFID Technology-based SLAM
100
90
Cart (Cartesian-Track Error)
Corrected
Uncorrected
80
70
Error [m]
60
50
40
30
20
10
1 2 3 4 5 6 7
Pedestrian ID
Figure 4.13: Result from the outdoor experiment: the average Cartesian error.
Decentralized SLAM Experiment
The last experiment was conducted in order to evaluate the decentralized version of
RFID-SLAM. This was carried out by simulating the incremental execution of trajectories
recorded during the outdoor experiment previously described. We executed all
720 possible sequences of six trajectories and computed the mean Cartesian error. After
the execution of each trajectory the graph learned by the according agent and the
graphs learned by passed RFIDs were computed. Therefore, the first pedestrian had no
advantage at all, the second pedestrian benefited from data propagated into RFIDs by
the first, and the third benefited from data propagated by the first and the second, and
so on. Note that graphs learned by the RFIDs do not necessarily consist of complete
trajectories, instead they contain the union of trajectories from the agents until they synchronized
this node. Figure 4.14 depicts the average Cartesian error after correcting the
trajectory of each pedestrian based on the graph generated from the union of all visited
RFIDs. The more pedestrians traveled through the environment, the better the pose estimate
of a single pedestrian can be improved. Hence, the last pedestrian was able to
achieve a nearly two times higher pose accuracy than the first one by closing loops that
were constructed by joining graphs from other pedestrians.
4.5.4 RFID-SLAM jointly with humans and robot
In another experiment we jointly corrected the odometry trajectories from a human and
a robot exploring the same area while detecting RFIDs. During this experiment, pedestrian
and robot performed pose tracking with the PDR method and slippage sensitive

4.5. Experimental results 89
55
50
Mean Error [m]
45
40
35
30
25
20
1 2 3 4 5 6
i-th pedestrian walking
Figure 4.14: Result from decentralized SLAM: Average Cartesian pose error of the
i-th pedestrian walking after his predecessors which left information locally in RFIDs.
Averages are computed from all possible sequences of six pedestrians.
odometry described in Chapter 3. For an area of approximately 900 m 2 , 10 RFIDs were
used. Table 4.1 depicts the average Cartesian error, the average cross-track error (XTE),
Cart. Err. [m] XTE Err. [m] ATE Err. [m]
Rob. Odo. 147.10 ± 36.85 139.59 ± 35.99 46.11 ± 20.69
Ped. Odo. 56.63 ± 24.38 44.22 ± 24.52 32.31 ± 15.82
Ped. Corr. 14.27 ± 12.7 8.40 ± 11.67 10.68 ± 10.84
Rob. Corr. 9.37 ± 9.90 5.57 ± 9.55 6.52 ± 9.23
Both Corr.. 5.64 ± 4.77 2.50 ± 4.23 4.33 ± 4.50
Table 4.1: Average positioning errors of odometry, single, and joint correction.
and average along-track error (ATE) of the original robot odometry (Rob. Odo.), the
original pedestrian odometry (Ped. Odo.), their single corrected trajectories (Ped. Corr,
Rob Corr.), and the jointly corrected trajectory (Both Corr.). The simultaneous correction
of both trajectories improved the accuracy significantly. Figure 4.15 (a) shows both
odometry trajectories compared to ground truth (blue line), and Figure 4.15 (b) shows
the corrected RFID graph (green line). The small squares indicate RFID observations.

90 Chapter 4. RFID Technology-based SLAM
(a)
(b)
Figure 4.15: Result from correcting trajectories from robot odometry (orange line) and
pedestrian odometry (red line) jointly: The corrected RFID graph (green line) lies close
to ground truth (blue line). Small squares indicate RFID observations.
4.6 Related work
During the last decade, methods for SLAM have been applied in many different scenarios,
for example in outdoor [Bailey, 2002, Ramos et al., 2007] and underwater environments
[Newman et al., 2003]. Inspired by the fundamental work of Smith et al. [Smith
et al., 1988], early work on SLAM was mainly based on the Extended Kalman Filter
(EKF) [Dissanayake et al., 2001] which updates the state vector after each measurement
in O ( n 2) . Based on the observation that landmark estimates are conditionally independent
given the robot’s pose, Montemerlo et al. introduced FastSLAM which reduces
the computational complexity of EKF-based SLAM to O (nk), where k is the number
of robot trajectories considered at the same time [Montemerlo et al., 2002]. The framework
has been further extended to using evidence grids [Hähnel et al., 2003, Grisetti
et al., 2005]. Stachniss et al. introduced a method for improving map quality gained
by FastSLAM during exploration by inducing actions on the robot for actively closing
loops, e.g. to cause the robot to re-enter already visited places [Stachniss et al., 2004].
Thrun et al. introduced an approach following the idea of representing uncertainty with
an information matrix instead of a covariance matrix [Thrun et al., 2004]. By exploiting
the sparsity of the information matrix, the algorithm, called Sparse Extended Information
Filter (SEIF), allows updates of the state vector in constant time. Another variant
of SLAM, the Treemap algorithm, has been introduced by Frese [Frese, 2006]. This
method divides a map into local regions and subregions and the landmarks of each re-

4.6. Related work 91
gion are stored at the according level of the tree hierarchy. Lu and Milios introduced
a method for globally optimizing robot trajectories by building a constraint graph from
LRF and odometry observations [Lu and Milios, 1997]. The method described in this
thesis is closely related to their work, but the modification enables efficient route graph
corrections by decomposing the problem into pose tracking, optimization, and interpolation.
In contrast to incrementally full state updates performed by EKF-based methods
after each observation, the decomposition reduces the computational requirements during
runtime to a minimum, thus allowing the efficient optimization of even large-scale
environments. Whereas existing methods typically rely on a high density of landmarks,
the RFID-based approach is tailored for very sparse landmark distributions with reliable
data association.
In connection with radio transmitters, the SLAM problem has mainly been addressed
as “range-only” SLAM [Kehagias et al., 2006, Djugash et al., 2005, Kurth et al., 2003,
Kantor and Singh, 2002] since the bearing of the radio signal cannot accurately be determined.
RFIDs have been successfully utilized already for localizing mobile robots
[Hähnel et al., 2004,Bohn and Mattern, 2004] and emergency responders [Kantor et al.,
2003,Miller et al., 2006]. Hähnel and colleagues [Hähnel et al., 2004] successfully utilized
Markov localization for localizing a mobile robot in an office environment. Their
approach deals with the problem of localization in a map previously learned from laser
range data and known RFID positions, whereas the work presented in this paper describes
a solution that performs RFID-based localization and mapping simultaneously
during exploration. Also sensor networks-based Markov localization for emergency
response has been studied [Kantor et al., 2003]. In this work, existing sensor nodes
in a building are utilized for both localization and computation of a temperature gradient
from local sensor node measurements. Bohn and colleagues examined localization
based on super-distributed RFID tag infrastructures [Bohn and Mattern, 2004]. In
their scenario tags are deployed beforehand in a highly redundant fashion over large
areas, e.g. densely integrated into a carpet. They outline the application of a vacuumcleaner
robot following these tags. Pedestrian dead reckoning has also been combined
with WLAN perception-based localization [Retscher and Kealy, 2006]. Miller and colleagues
examined the usability of various RFID systems for the localization of first
responders in different building classes [Miller et al., 2006]. During their experiments,
persons were tracked with a Dead Reckoning Module (DRM) while walking through a
building. They showed that the trajectories can be improved by utilizing the positions
of RFID tags detected in the building. While these map improvements were carried
out with only local consistency, the approach presented in this work yields a globally
consistent map improvement.
Efficient solutions for decentralized SLAM have been introduced in the past [Reece
and Roberts, 2005, Nettleton et al., 2003, Kim et al., 2004]. These approaches assume
that at least low-bandwidth radio communication between the agents is possible. They
solve the double counting problem, e.g. the multiple propagation of equal information

92 Chapter 4. RFID Technology-based SLAM
via different paths, by probabilistic methods, such as Bounded Covariance Inflation
(BCI) [Reece and Roberts, 2005], and Covariance Intersection (CI) [Nettleton et al.,
2003]. The decentralized method proposed in this work shows that this problem can be
solved also by managing estimates locally in the memory of RFIDs.
4.7 Conclusion
RFID-SLAM allows the efficient generation of globally consistent maps, even if the
density of landmarks is comparably low. For example, the method corrected an outdoor
large-scale map within a few seconds from odometry data and RFID perceptions
only. This was partially achieved due to reliable pose tracking based on slippage sensitive
odometry, but also due to the data association solved by RFIDs. Solving data
association by RFIDs allows to speed-up the route graph corrections by decomposing
the problem into optimization and interpolation. Furthermore, a novel method for distributed
SLAM by exploiting the memory capacity of RFIDs, has been contributed.
The result shows that sharing information between single agents allows to globally optimize
their individual paths, even without explicit need for communication. This has
the advantage that, particularly in the context of disaster response, this method can also
be applied if communication is disturbed by building debris and radiation. If communication
is available, the process can be further accelerated by exchanging graph data
between the agents directly.
Moreover, the introduced method allows to apply SLAM without requiring pedestrians
and robots to perform loops while executing their primary task. Due to the joining
of routes via RFID connection points, loops automatically emerge, which is a necessary
requirement if applying SLAM in the real-world, particularly when humans are
involved in USAR (Urban Search And Rescue) situations. Finally, the joint correction
of human and robot trajectories, has been demonstrated. The result shows that sharing
information between independent agents, i.e. humans and robots, allows to correct their
individual paths globally.
Besides, RFID-SLAM offers many advantages, particularly in a disaster response
scenario. One practical advantage is that humans can be integrated easily into the search
since the exchange of maps can be carried out via the memory of RFIDs [Kleiner and
Sun, 2007]. Furthermore, they can communicate with RFIDs by a PDA and leave behind
information related to the search or to victims. The idea of labeling locations with
information that is important to the rescue task, has already been applied in practice.
During the disaster relief in New Orleans in 2005, rescue task forces marked buildings
with information concerning, for example, hazardous materials or victims inside the
buildings [FEMA, 2003]. The RFID-based marking of locations is a straight forward
extension of this concept.
In future work, we will evaluate RFID technology operating in the UHF frequency

4.7. Conclusion 93
domain, which allows reading and writing within distances of meters. As we have
already demonstrated in former work [Ziparo et al., 2007b], the combination of an
RFID route graph representation with local mapping opens the door to efficient large
scale exploration and mapping. Additionally, we will deal with the problem of building
globally consistent elevation maps, as described in Chapter 5, by utilizing RFID
technology-based route graph optimization for loop-closure.

5 Real-time Elevation Mapping
5.1 Introduction
The motivation for the elevation mapping method contributed in this chapter is to provide
a basis for enabling robots to continuously plan and execute skills on rough terrain.
Therefore, the method has to be computational efficient and capable of building elevation
maps that are sufficiently accurate for structure classification and behavior planning,
as we have already demonstrated in former work [Dornhege and Kleiner, 2007b].
Building globally consistent elevation maps, e.g. by loop closure, is computationally
hard in large-scale environments since it requires to keep the whole map in memory.
Therefore, the goal of the proposed method is not to build globally consistent maps, but
maps that are locally consistent in the vicinity of the robot for continuous planning and
navigation. In a further processing step, e.g. offline, a global map can be generated by
merging locally consistent map patches according to a globally consistent route graph,
as for example, computed by the RFID-SLAM method introduced in Chapter 4. Nevertheless,
in the following we will denote poses as global in order to distinguish between
the local and global coordinate frame of the robot.
Elevation mapping is carried out with a Kalman filter-based approach by integrating
range measurements from a downwards tilted laser range finder. The map is incrementally
build in real-time while the mobile robot explores an uneven surface. The method
tracks the three-dimensional pose of the robot by integrating the robot’s orientation,
and the two-dimensional pose generated from visual odometry and scan matching (see
Chapter 3). Furthermore, the three-dimensional pose is updated from height observations
that have been registered on the map. Given the three-dimensional pose, the height
value of each map cell is estimated by a Kalman filter that integrates readings from a
downwards tilted LRF. Due to the integration of the three-dimensional pose, the method
allows to create elevation maps while the robot traverses rough terrain, as for example
driving over ramps and stairs.
The proposed approach was extensively evaluated in USAR test arenas designed by
the National Institute of Standards and Technology (NIST) [Jacoff et al., 2001]. The
results show that the method allows to reliably generate elevation maps in real-time
while the robot explores an unknown area.
The remainder of this chapter is structured as follows. In Section 5.2 state-of-the-art
techniques for mapping are described. In Section 5.3 the method for real-time elevation
95

96 Chapter 5. Real-time Elevation Mapping
mapping is proposed, and in Section 5.4 empirical results from mapping experiments
are presented. Finally, in Section 5.5 an overview on related work is given, and in
Section 5.6 the conclusion presented.
5.2 Conventional techniques
5.2.1 Occupancy grid maps
Occupancy grid maps are a data structure for fusing multiple sensor measurements
into a discretized metric representation. They were introduced by Elfes and Moravec
[Moravec and Elfes, 1985, Elfes, 1989], and have mainly been utilized for integrating
range measurements from sonar and laser sensors for building a two-dimensional map
of the environment. Basically, the representation partitions the space around the robot
into equally sized rectangular cells. Each cell can take on values between 0 and 1, describing
the probability that the corresponding location in the real world is occupied by
an obstacle. Figure 5.1 depicts an occupancy grid generated in an office-like environment.
The darker a cell, the higher the probability of occupancy. It is usually assumed
that unexplored regions have an occupancy probability of 0.5, which corresponds to the
large gray region in the figure. Elfes and Moravec developed an update procedure for
(a)
(b)
Figure 5.1: (a) Occupancy grid map of an office-like environment, incrementally
generated from laser range data (blue). (b) The robot plans to the next frontier cell
(green) [Yamauchi, 1997].
computing for each map cell c xy the occupancy probability P ( occ xy | o 1:t , l 1:t
)
, given past
sensor observations o 1:t = o 1 , · · · , o t , and robot poses l 1:t = l 1 , · · · , l t . This is carried out
by applying Bayes’ rule under the assumption of independence between single sensor
readings. Consequently, the occupancy probability of cell c xy can be updated according

5.2. Conventional techniques 97
to the following procedure:
P ( )
occ xy | o 1:t , l 1:t =
⎡
1 − ⎢⎣ 1 + P ( )
occ xy | o t , l t
1 − P ( ) 1 − P ( )
occ xy
occ xy | o t , l t P ( P ( ) ⎤
occ xy | o 1:t−1 , l 1:t−1
)
occ xy 1 − P ( ) ⎥⎦
occ xy | o 1:t−1 , l 1:t−1
−1
(5.1)
The procedure can be applied incrementally each time new sensor readings are available.
In order to avoid numerical instabilities for probabilities close to zero or one,
Equation 5.2.1 can be transformed to the log odds representation [Thrun et al., 2005].
Note that in static environments the prior probability P ( occ xy
)
is typically assumed to
be 0.5, thus the corresponding term in Equation 5.2.1 can be omitted. The update requires
the global pose of the robot, which can be computed by methods described in
Chapter 3 and Chapter 4. Furthermore, the occupancy probability P ( occ xy | o t , l t
)
is
computed based on a inverse sensor model, which is specific for the utilized type of
sensor. In case of laser range finders, the sensor model is simply defined by a piecewise
linear function:
P ( occ xy | o t , l t
)
=
⎧⎪⎨⎪⎩
P prior
P occ
P f ree
if c xy is not covered by the beam
if c xy is at the end of the beam
if c xy is before the end of the beam,
(5.2)
where typically P prior = 0.5, P occ = 0.8, P f ree = 0.2.
It can be seen in Figure 5.1 that the incremental update of the map reveals the structure
of the environment, e.g. doors and hallways. The representation can be utilized efficiently
for exploration and planning. Frontier cell-based exploration [Yamauchi, 1997]
is a method that selects exploration targets based on the occupancy values of each cell
and its next neighbors. Frontier cells are unoccupied cells and neighbor at least one cell
that is unexplored. They are extracted from the grid map and taken as target locations
for a trajectory planner. Figure 5.1 (b) depicts the generated frontier cells (green), and
a path (blue) planned to reach one of these cells. Planning can be carried out by either
applying value iteration or A* planning [Russell and Norvig, 2003] on the grid map,
after growing obstacles on the grid with a Gaussian kernel in order to account for the
size of the robot.
The described method has been widely used in office-like environments. In unstructured
environments, a three-dimensional representation is needed since the surface
might be uneven, and obstacles can be located at different levels of height. Moravec
extended the approach for building occupancy grids in three-dimensions from stereo
vision [Moravec, 1996]. However, since the computational requirements significantly
increase in three-dimensions, the method does not scale in large environments.

98 Chapter 5. Real-time Elevation Mapping
5.3 Building elevation maps in real-time
An elevation map is represented by a two-dimensional array storing for each global
location (x g , y g ) a height value h with variance σ 2 h
. In order to determine the height
for each location, endpoints from the LRF readings are transformed from robot-relative
distance measurements to global locations, with respect to the robot’s global pose, and
the pitch (tilt) angle of the LRF (see Figure 5.2).
This section is structured as follows. In Section 5.3.1 the update of single cell values
relative to the location of the robot is described, in Section 5.3.2 the filtering of the map
with a convolution kernel is shown, and in Section 5.3.3 an algorithm for the estimation
of the robot’s three-dimensional pose from dead reckoning and map observations is
proposed.
Figure 5.2: Transforming range measurements to height values.
5.3.1 Single cell update from sensor readings
Our goal is to determine the height estimate for a single cell of the elevation map with
a Kalman filter [Maybeck, 1979], given all height observations of this cell in the past.
We model height observations z t by a Gaussian distribution N ( )
z t , σ 2 z t , as well as the
current estimate N
(ĥ (t) , σ
2
ĥ(t)
)
of each height value. Note that the height of cells cannot
be observed directly, and thus has to be computed from the measured distance d and
LRF pitch angle α. Measurements from the LRF are mainly noisy due to two error
sources. First, the returned distance depends on the reflection property of the material,
ranging from very good reflections, e.g. white sheet of paper, to nearly no reflections,
e.g. black sheet of paper. Second, in our specific setting, the robot acquires scans while
navigating on rough terrain. This will lead to strong vibrations on the LRF, causing an
oscillation of the laser around the servo-controlled pitch angle. Consequently, we represent
measurements from the LRF by two normal distributions, one for the measured
distance N(µ d , σ d ), and one for the pitch angle N(µ α , σ α ).
The measurements from the LRF are transformed to robot-relative locations (x r , y r ).
First, we compute the relative distance d x and the height z of each measurement accord-

5.3. Building elevation maps in real-time 99
ing to the following equation (see Figure 5.2):
( ) ( ) (
dx d
= F
z dα =
α
d cos α
h R − d sin α
)
, (5.3)
where h R denotes the height of the LRF mounted on the robot. Second, from distance
d x and the horizontal angle β of the laser beam, the relative cell location (x r , y r ) of each
cell can be calculated by:
x r = d x cos β (5.4)
y r = d x sin β (5.5)
Equation 5.3 can be utilized for computing the normal distributed distance N(µ dx , σ dx ),
and height N(µ z , σ z ), respectively. However, since this transformation is non-linear, F dα
has to be linearized by a Taylor series at µ dx , µ z :
( ) ( )
µdx d
= F
µ dα (5.6)
z α
Σ dx z = ∇F dα Σ dα ∇F T dα (5.7)
( )
cos α −d sin α
with ∇F dα =
(5.8)
− sin α −d cos α
( ) σ
2
and Σ dα =
d
0
0 σ 2 . (5.9)
α
Then, the height estimate ĥ can be updated from observation z t , taken at time t, with the
following Kalman filter:
ĥ (t) =
1
σ 2 z t
+ σ 2 ĥ(t−1)
(
σ
2
zt
ĥ (t − 1) + σ 2 ĥ(t−1) z t)
σ 2 ĥ(t) = 1
1
σ 2 ĥ(t−1)
+ 1
σ 2 zt
(5.10)
. (5.11)
Equation 5.10 cannot be applied if the tilted LRF scans vertical structures since they
lead to different height measurements for the same map location. For example, close
to a wall the robot measures the upper part, far away from the wall the robot measures
the lower part. We restrict the application of the Kalman Filter by the Mahalanobis
distance. If the Mahalanobis distance between the estimate and the new observation is
below a threshold c, the observation is considered to be taken from the same height. We
use c = 1, which has the effect that all observations with a distance to the estimate that
is below the standard deviation σĥ, are merged. Furthermore, we are mainly interested

100 Chapter 5. Real-time Elevation Mapping
in the maximum height of a cell, since this is exactly what elevation maps represent.
These constraints lead to the following update rules for cell height values:
⎧
z t
⎪⎨
ĥ (t) = ĥ (t − 1)
1 ⎪⎩
σ 2 zt +σ2 ĥ(t−1)
and variance σ 2 ĥ(t) with:
(
)
σ 2 z t
ĥ (t − 1) + σ 2 z ĥ(t−1)
t
if z t > ĥ (t) ∧ d M
(
zt , ĥ (t) ) > c
if z t < ĥ (t) ∧ d M
(
zt , ĥ (t) ) > c
else,
(5.12)
σ 2 ĥ(t) = ⎧⎪⎨⎪⎩
σ 2 z t
σ 2 ĥ(t−1)
1
1
σ 2 + 1
σ 2 ĥ(t−1)
zt
if z t > ĥ (t) ∧ d M
(
zt , ĥ (t) ) > c
(
if z t < ĥ (t) ∧ d M zt , ĥ (t) ) > c
else,
(5.13)
where d M denotes the Mahalanobis distance, defined by:
d M
(
zt , ĥ(t) ) =
√(
zt − ĥ(t) )
σ 2 ĥ(t)
2
. (5.14)
The cell update introduced so far assumes perfect information on the global pose of
the robot. However, since we integrate measurements from the robot while moving in
the environment in real-time without loop-closure ∗ , we have to account for positioning
errors from pose tracking that do accumulate over time. Continuous Kalman updates
without regarding pose uncertainty due to robot motion between update steps will successively
reduce the cell’s variance leading to variance estimates which highly underestimate
the actual uncertainty about a cell’s height. Thus we increase the variance in the
Kalman propagation step based on the robot’s motion to reflect the true uncertainty in
the variance estimate. The height itself is not changed as the old estimate still gives the
best possible estimation for a cell’s height. We assume that the positioning error grows
linearly with the accumulated distance and angle traveled. Hence, observations taken
in the past loose significance with the distance the robot traveled after they were made:
ĥ(t) = ĥ(k) (5.15)
σ 2 ĥ(t) = σ2ˆk(t) + σ2 d d(t − k) + σ2 αα(t − k), (5.16)
where t denotes the current time, k denotes the time of the last height measurement at
the same location, d(t−k) and α(t−k) denotes the traveled distance and angle within the
∗ Note that global localization errors can be reduced by loop-closure, i.e. by re-computing the elevation
map based on the corrected trajectory, which, however, can usually not be applied in real-time.

5.3. Building elevation maps in real-time 101
time interval t − k, and σ 2 d , σ2 α are variances that have to be determined experimentally
according to the utilized pose tracker. Since it would be computationally expensive to
update the variances of all grid cells each time the robot moves, updates according to
Equation 5.16 are only carried out on variances before they are utilized for a Kalman
update with a new observation. The traveled distances can be efficiently generated by
maintaining the integral functions I d (t) and I α (t) that provided the accumulated distance
and angle for each discrete time step t, respectively. Then, for example, d(k − t) can be
calculated by I d (k) − I d (t). The integrals are represented by a table indexed by time t
with a fixed discretization, e.g. ∆t = 1 s.
5.3.2 Map filtering with a convolution kernel
The limited resolution of the LRF occasionally leads to missing data in the elevation
map, e.g. conspicuous by surface holes. Furthermore, the effect of “mixed pixels” might
lead to phantom peaks within the elevation map [Ye and Borenstein, 2003], which
frequently happens if the laser beam hits edges of objects and the returned distance
measure is a mixture of the distance to the object and the distance to the background.
Therefore, the successively integrated elevation map has to be filtered.
In computer vision, filtering with a convolution kernel is implemented by the convolution
of an input image with a kernel in the spatial domain, i.e. each pixel in the
filtered image is replaced by the weighted sum of the pixels in the filter window. The
effect is that noise is suppressed and the edges in the image are blurred at the same
time. We apply the same technique on the elevation map in order to reduce the errors
described above. Hence, we define a convolution kernel of the size of 3 × 3 cells, where
each value is weighted by its certainty and distance to the center of the kernel. Let
h(x + i, y + j) denote a height value relative to the kernel center at map location (x, y),
with i, j ∈ {−1, 0, 1}. Then, the weight for each value is calculated as follows:
⎧
1
if |i| + | j| = 0
σ
⎪⎨
2 h(x+i,y+ j)
1
w i, j =
if |i| + | j| = 1
2σ 2 (5.17)
h(x+i,y+ j)
1 ⎪⎩ if |i| + | j| = 2
4σ 2 h(x+i,y+ j)
Consequently, the filtered elevation map h f can be calculated by:
where C = ∑ w i, j .
h f (x, y) = 1 ∑
h(x + i, y + j)w i, j , (5.18)
C
i, j

102 Chapter 5. Real-time Elevation Mapping
5.3.3 Estimation of the three-dimensional pose
So far the incremental procedure for updating elevation map cells relative to the coordinate
frame of the robot has been shown. In order to update map cells globally, the
three-dimensional pose of the robot has to be considered, which is described by the vector
l = (x, y, h, θ, φ, ψ) T , where θ denotes the yaw angle, φ denotes the pitch angle, and
ψ denotes the roll angle, respectively. We assume that IMU measurements of the three
orientation angles are given with known variance. The position (x, y, h) is estimated by
dead reckoning, which is based on the pitch angle and traveled distance measured by
visual odometry (see Chapter 3) and scan matching. However, since these measurements
are estimating the relative displacement δ with respect to the three-dimensional
surface, δ has to be projected onto the plane, as depicted in Figure 5.3. Given the input
Figure 5.3: Dead reckoning of the projected Cartesian position (x p , y p , h p ) from yaw
angle θ, pitch angle φ, and traveled distance δ.
u = (θ, φ, δ) T represented by the Gaussian distribution N(µ u , σ u ), the projected position
l = (x p , y p , h p ) T , represented by the Gaussian distribution N (µ l , Σ l ), can be calculated
as follows:
⎛
⎜⎝
x p t
y p t
h p t
⎞
⎟⎠ = F lu
⎛
x p ⎞
t−1
y p t−1
h t−1
φ
θ ⎜⎝ ⎟⎠
δ
⎛
= ⎜⎝
x p t−1
y p t−1
+ δ cos θ cos φ
+ δ sin θ cos φ
⎞⎟⎠
h p t−1 + δ sin φ
(5.19)
Σ lu = ∇F lu Σ lu ∇F T lu (5.20)
Σ lu = ∇F l Σ l ∇F T l + ∇F u Σ u ∇F T u , (5.21)

5.4. Experimental results 103
where
⎛ ⎞
1 0 0
∇F l = 0 1 0 ⎜⎝ ⎟⎠ , (5.22)
0 0 1
⎛
⎞
−δ cos θ sin φ −δ sin θ cos φ cos θ cos φ
∇F u = −δ sin θ sin φ δ cos θ cos φ sin θ cos φ
⎜⎝
⎟⎠ , (5.23)
δ cos φ 0 sin φ
⎛
σ 2 ⎞
φ
0 0
∇F u = Σ u = 0 σ ⎜⎝
2 θ
0 ⎟⎠ , (5.24)
0 0 σ 2 δ
⎛
σ 2 x p σ2 x p y p σ2 x p h p
σ 2 x p y p σ2 y p σ2 y p h p
Σ l =
⎜⎝
σ 2 x p h
σ 2 p y p h
σ
⎞⎟⎠ . (5.25)
2 p h p
Equation 5.19 allows to predict the current height of the robot. However, due to the accumulation
of errors, the accuracy of the height estimate will decrease continuously.
Therefore, it is necessary to update this estimate)
from direct observation. For this
purpose, we utilize the height estimate
(ĥ (t) , σ
2
at the robot’s position from Equation
5.12 and 5.13, ) respectively. Then, the new estimate can be calculated by Kalman-
ĥ(t)
fusing
(ĥ (t) , σ
2
with the predicted height estimate ( )
h p , σ 2 ĥ(t)
h analogous to Equation
5.10.
p
The global location (x g , y g ) of a measurement, i.e. the elevation map cell for which
the height estimate ĥ (t) will be updated, can be calculated straightforwardly by:
x g = x r + x p (5.26)
y g = y r + y p (5.27)
5.4 Experimental results
Elevation mapping was evaluated on the Lurker robot (see Chapter 2), which is capable
to autonomously overcome rough terrain containing ramps and rolls. The system was
successfully demonstrated during the RoboCup Rescue autonomy competition in 2006,
where the robot won the first prize. The testing arena, which was used for the exper-

104 Chapter 5. Real-time Elevation Mapping
iments presented in this section, was designed by NIST during the Rescue Robotics
Camp 2006 with the same degree of difficulty as presented at RoboCup‘06, e.g. also
containing rolls and ramps. During all experiments, the robot was equipped with an
IMU sensor, a side camera for visual odometry, and two LRFs, one for scan matching
and one for elevation mapping. The latter sensor was tilted downwards by 35 ◦ .
Figure 5.4 depicts the Kalman filter-based pose estimation of the robot, as described
in Section 5.3.3. For this experiment, conditions were made intentionally harder. Map
smoothing was turned off, which had the effect that missing data, due to a limited resolution
of 2D scans, lead to significant holes on the surface of the map. Furthermore,
we added a constant error of −2 ◦ to pitch angle measurements of the IMU. As shown
in Figure 5.4, the Kalman filter is able to cope with such errors, and finally produced a
trajectory close to ground truth (indicated by the gray surface).
(a)
(b)
Figure 5.4: Evaluation of the Kalman filter for tracking the robot’s height. (a) Height
values predicted from the IMU (red line) are merged with height values taken from the
generated map (blue line). Errors from inaccuracies in the map, as well as a simulated
continuously drift error of the IMU sensor are successfully reduced (green line). (b)
Merged trajectory compared to ground truth (gray ramp).
In order to quantitatively evaluate the visual odometry support for elevation mapping,
we recorded hight estimates while the robot was autonomously exploring an area,
and finally climbing-up an open ramp. Figure 5.5 shows the results of the Kalman
filter-based height estimation (blue line with crosses) in comparison to the manually
measured ground truth (black line with triangles). As shown in Figure 5.5, the Kalman
filter computes continuously an estimate close to ground truth, which stays consistently
within the expected covariance bounds.
Figure 5.6 shows the final result from applying the proposed elevation mapper during
the Rescue Robotics camp in Rome 2006. Figure 5.6 (a) depicts an overview on the
arena, and Figure 5.6 (b) shows the calculated height values. The height of each cell is
indicated by a gray value, the darker the cell, the bigger the elevation. Figure 5.6 (c)
depicts the variance of each height cell, going from pink (hight variance) to yellow (low
variance) The current position of the robot is indicated by a blue circle in the lower left
corner. The further away cell updates on the robot’s trajectory, the lower their variance

5.5. Related work 105
1000
800
Height Estimation
Ground Truth
3σ Bound around Ground Truth
600
Height [mm]
400
200
0
-200
-400
0 50 100 150 200 250
Time [s]
Figure 5.5: Quantitative evaluation of the Kalman-filtered height estimation. During
this experiment the robot started exploration on the floor and then climbed a ramp at
around 220 s. The graph shows the ground truth (black line with triangles) in comparison
to the height estimation by the Kalman filter (blue line with crosses).
(see Section 5.3.1). Figure 5.6 (d) shows a 3D visualization of the generated elevation
map. Structures, such as the long ramp at the end of the robot’s trajectory, and the stairs,
can clearly be identified. We measured on an AMD64X2 3800+ a total integration time
(without map smoothing) of 1.88 ± 0.47 ms for a scan measurement with 683 beams,
including 0.09 ± 0.01 ms for the 3D pose estimation. Map smoothing has generally the
time complexity of O ( N 2 M ) , where N is the number of rows and columns of the map
and M the size of the kernel. We measured on the same architecture 34.79 ± 14.84 ms
for smoothing a map with N = 300 and M = 3. However, this can be improved
significantly during runtime by only smoothing recently modified map cells and their
immediate neighbors within distance M.
5.5 Related work
Elevation maps are indispensable, particularly for robots operating in unstructured environments.
They have been utilized on wheeled robot platforms [Pfaff et al., 2007a,Wolf
et al., 2005], on walking machines [Krotkov and Hoffman, 1994,Gassmann et al., 2003]
and on car-like vehicles [Thrun et al., 2006,Ye and Borenstein, 2003,Pfaff et al., 2007b].

106 Chapter 5. Real-time Elevation Mapping
(a)
(b)
(c)
(d)
Figure 5.6: Elevation mapping during the Rescue Robotics Camp 2006 in Rome: (a)
The arena build-up by NIST, (b) the corresponding digital elevation model (DEM) build
by the Lurker robot, going from white (low altitude) to black (hight altitude), (c) the
variances of each height value according to the robots position (lower left corner), going
from pink (hight variance) to yellow (low variance), (d) the 3D perspective rotated 180 ◦ .
These methods differ in the way how range data is acquired. If data is acquired from a
3D scan [Pfaff et al., 2007a, Krotkov and Hoffman, 1994], it usually suffices to employ
standard error models, which reflect uncertainty from the measured beam length. Data
acquired from a 2D LRF, e.g. tilted downwards, requires more sophisticated error models,
such as the compensation of pose uncertainty [Thrun et al., 2006], and handling of
missing data by map smoothing [Ye and Borenstein, 2003].
Other researchers work focused on the building of maps from 3D scans by registering
each scan on the existing map with the Iterative Closest Points (ICP) algorithm [Surmann
et al., 2003]. In Surmann’s setting, the robot acquires a 3D scan, plans and navigates
based on this scan with a next best view planner, and stops again for acquiring the
next scan from another position. They extended their work with a method for generat-

5.6. Conclusion 107
ing Prolog-based semantic descriptions of the 3D point clouds, which they utilized for
optimizing the 3D model by Prolog unification [Nüchter et al., 2003]. Generally, full
3D data processing is still only limitedly applicable in real-time. The acquisition of 3D
data, as well as 3D data registration, is still time consuming, thus requiring interruptions
of continuous navigation. Furthermore, the sporadic data acquisition makes it difficult
to apply the method in dynamic environments.
In contrast to previous work, the proposed approach deals with the problem of building
elevation maps in real-time, allowing the robot continuous planning and navigation.
Furthermore, the assumption that the robot has to be situated on a flat surface while
mapping rough terrain has been relaxed.
5.6 Conclusion
We have contributed a solution for the real-time building of elevation maps from LRF
data. It has been demonstrated that the generated elevation maps provide a good tradeoff
between computational efficiency and structural representation. They can reliably
be generated in real-time while the robot is continuously in motion. This has partially
been achieved by the introduced method for visual odometry, which significantly improved
the accuracy of scan matching. As former work has shown, the generated elevation
maps can be utilized for structure classification, and the planning of skill execution
[Dornhege and Kleiner, 2007b], thus enabling a mobile robot to truly autonomously
explore rough terrain.
To explore terrain based on a local world model might be sufficient for a range of
tasks. However, many real-world situations require the robot to continuously build a
global representation of the world in order to perform tasks globally efficient. For example,
the problem of searching victims after a disaster requires a global representation of
a large-scale environment. To represent such environments by grid map-like data structures
becomes intractable if the map size increases. Therefore, in future work, we will
deal with the problem of building globally consistent elevation maps by utilizing RFID
technology-based route graph optimization for loop-closure (see Chapter 4). Here the
basic idea is to anchor locally generated elevation maps in the corrected graph. These
map patches can then be loaded into memory if they are within range of the robots current
location. Furthermore, the saved data can be utilized offline for generating a global
elevation map representation of the environment.

6 RFID Technology-based
Multi-Robot Exploration
6.1 Introduction
During rescue operations for disaster mitigation cooperation is a must [Jennings et al.,
1997]. In general the problem is not solvable by a single agent, but a heterogeneous
team is needed that dynamically combines individual capabilities in order to solve the
task. This requirement is due to the structural diversity of disaster areas, various sources
of human presence that sensors can perceive, and to the necessity of quickly and reliably
examining the targeted regions.
In the urban search and rescue context, there exists no standard robot platform that is
capable of solving all kinds of challenges demanded by the environment. For example,
there are places only reachable by climbing robots, spots only accessible through small
openings, and regions only observable by aerial vehicles. Multi-robot teams do not
only offer the possibility to field such diverse capabilities, they also exhibit increased
robustness due to redundancy, and superior performance due to parallel task execution
[Arai et al., 2002]. This latter aspect is as important as the first one, since the time
to complete a rescue mission is literally of vital importance. Therefore, to approach
the task of designing a multi-robot system for rescue from single-robot solutions is
prone to yield suboptimal solutions. The joint performance of a team depends on the
right mixture and coordination of the individual robot capabilities, and thus has to be
designed as a whole. One fundamental problem in this context is the coordination of a
team during an exploration task.
To coordinate a team of robots for exploration is a challenging problem, particularly
in large-scale areas, as for example the devastated area after a disaster. This problem can
be generally decomposed into task assignment and multi-agent path planning. Whereas
task assignment and multi-agent path planning were both intensively studied as separate
problems in the past, there has been only little attention on solving both of them at once,
particularly if large robot teams are involved. This is mainly due to the fact that the
joint state space of the planning problem grows enormously with the number of robots.
However, particularly in destroyed environments where robots have to overcome narrow
passages and obstacles, path coordination is essential in order to avoid collisions and
deadlocks.
109

110 Chapter 6. RFID Technology-based Multi-Robot Exploration
In this chapter a novel deliberative approach that reduces significantly the size of
the search space by utilizing RFID tags as coordination points, while solving the problem
of task assignment and path planning simultaneously, is contributed. The method
utilizes the same RFID infrastructure as RFID-SLAM (Chapter 4), requiring robots to
deploy RFID tags autonomously in the environment for building a network of reachable
locations. Hence, global path planning is carried out on a graph topology, which is
computationally cheaper than planning on global grid maps, as it is usually the case.
Furthermore, a local search method that uses the memory of RFIDs for labeling visited
places in the field is contributed. Each RFID holds a set of poses that have been
visited by robots in the vicinity of the tag. Since data exchange is carried out via the
memory of RFIDs, the method can also be applied if radio communication fails completely.
Additionally, the local approach has the advantage that computational costs do
not grow with the number of robots participating in the search. Note that the correct
association of visited poses to RFID locations requires to determine detection range and
bearing from signal strength and antenna orientation, respectively.
The implemented system can be considered as two-layered, consisting of a local
mechanism and a global mechanism. The global mechanism monitors the local exploration,
and possibly restarts it at different locations, if the overall performance can be
improved. This is carried out by assigning new target locations and generating a global
multi-robot plan to reach them. Global planning is carried out in configuration timespace
on the RFID network, and genetic sequence selection is applied as a global task
assignment strategy.
For the purpose of navigation and obstacle avoidance, robots plan their trajectories
with A* [Russell and Norvig, 2003] on a dynamically updated grid map (see Section
5.2), which is limited in size and shifted according to robot motion. Due to the
limited size, grid updates from laser range data and navigation planning can be achieved
without significant computational costs.
The proposed methods were evaluated in the USARSim [Balakirsky et al., 2006]
simulation environment, which serves as basis for the Virtual Robots competition at
RoboCup [Balakirsky et al., 2007], where our team one the first prize of the competition.
The results show that the RFID tag-based exploration works for large robot teams,
particularly under limited computational resources.
The remainder of this chapter is organized as follows. The RFID technology-based
local and global exploration method and their application to rescue missions is discussed
in Section 6.2 and Section 6.4, respectively. In Section 6.3 a more detailed
description of the topological RFID map generated by the robots during exploration is
provided. In Section 6.5 results from experiments are shown and in Section 6.6 related
approaches are discussed. Finally, in Section 6.7 the conclusion is presented.

6.2. Coordinated local exploration 111
6.2 Coordinated local exploration
In this section a novel coordination mechanism [Kleiner et al., 2006c, Ziparo et al.,
2007a] is contributed, which allows robots to explore an environment with low computational
overhead and communication constraints. In particular, the computational costs
do not increase with the number of robots. The key idea is that the robots plan their path
and explore the area based on a local view of the environment. Team coordination is
realized through the use of indirect communication via RFID tags. In Section 6.2.1
the method for efficient navigation based on a local world view is presented, and in
Section 6.2.2 a novel coordination mechanism based on target selection and indirect
communication is introduced.
6.2.1 Navigation
To efficiently and reactively navigate, robots continuously plan their trajectories based
on a local representation of the environment, which is maintained within an occupancy
grid [Moravec, 1988] limited in size. For example, in our implementation, the local
map was limited to a four meter side square with 40 mm resolution. The occupancy
grid is shifted according to wheel odometry and IMU data, and continuously updated
from measurements of the LRF. This allows to overcome the accumulation of odometry
errors, while having some memory of the past. The exploration process periodically
selects targets, as shown in the following section, and continuously performs A*
search [Russell and Norvig, 2003] on the occupancy grid in order to generate plans for
reaching them. The continuous re-planning allows the robot to reactively avoid newly
perceived obstacles, and to deal with unforeseen situations caused by errors during path
following. A* planning is implemented with the Euclidean distance heuristic and the
state expansion selects all neighbors of a cell that are not obstructed (i.e. have an occupancy
value lower than a given threshold). The cost function c takes the length of the
path and its distance to obstacles into account:
c (s i+1 ) = c (s i ) + d (s i+1 , s i ) ∗ (1 + α ∗ occ (s i+1 )) (6.1)
where occ (s) is the probability of occupancy of grid cell s, d (.) is the Manhattan distance,
and α is a factor for varying the cost for passing nearby obstacles. Furthermore,
the grid is convoluted with a Gaussian kernel, forcing plans to be within a minimum
distance to nearby obstacles.
While navigating in the environment, robots maintain a Local RFID Set (LRS), which
contains all perceived RFIDs that are within range of the local occupancy grid. RFIDs
which are out of the antenna’s detection range, but still within the range of the grid,
are shifted according to the motion of the robot. Additionally, the occupancy grid is
augmented with RFIDs from a topological RFID graph representation, which will be

112 Chapter 6. RFID Technology-based Multi-Robot Exploration
further described in Section 6.4, facilitating the processing of nearby RFIDs that have
not been observed directly. On the basis of this information, new RFIDs are released in
the environment in order to maintain a predefined density of RFID tags (in our implementation
we took care of having the RFIDs at one meter distance from each other).
RFIDs that are within the local occupancy grid are utilized for avoiding collisions
between the robots. Each robot tracks its own pose by integrating measurements from
the wheel odometry and the IMU sensor with an Extended Kalman filter (EKF). As
commonly known, the accuracy of this estimate decreases due to the accumulation of
positioning errors, which can, for example, be prevented by performing data association,
as shown in Chapter 4. However, since the proposed approach aims at a computational
efficient implementation, pose estimates are not globally improved during the
multi-robot exploration. Instead, local displacements between robots are synchronized
via RFID tags. As will be shown, this method leads to pose estimates that are sufficiently
accurate for collision-free navigation. Note that after the execution, the global
map can still be improved offline, for example, by methods described in Chapter 4.
If two robots detected the same RFID tag in the past, the estimates of their mutual
displacement d R1 R 2
≈ l R1 − l R1 were synchronized by utilizing their local pose estimates
at this RFID: Let l R1 (t 1 ) and l R2 (t 2 ) denote the individual pose estimates of robot R 1 and
R 2 while detecting the same RFID tag at time t 1 and time t 2 , respectively. Then, the
new displacement between both robots can be calculated by d R1 R 2
= l R1 (t 1 ) − l R1 (t 2 ).
Furthermore, each robot can estimate poses within the reference frame of other robots
by utilizing the latest displacement and the individual pose estimate of the other robot
at time t. For example, R 2 ’s pose estimate of R 1 is given by: ˆl R1 (t) = l R1 (t) − d R1 R 2
.
Note that this procedure assumes the existence of a synchronized clock and requires the
robots to keep their trajectory in memory.
Known locations of other robots can be utilized on the planning level for avoiding
collisions between them. This is carried out by adding an extra cost to occupancy grid
cells that are close to these locations. If a robot detects that a teammate with a higher
priority is close, it stops until the other has moved out of the way. The simplest choice
for a priority scheme is to assign a ranking of the robots at start-up time, e.g. to assign to
each robot a unique ID. However, this method might fail if, for example, the robot with
the highest priority is unable to move first since it is blocked by any other robot with a
lower priority. There exist analytical solutions to this problem, as for example described
by Jaeger and Nebel [Jaeger and Nebel, 2001], which require as input a dependency
graph consisting of one node for each robot and one arc for each blockage between
two robots. Since it is hard to decide from sensor data, such as vision or laser range
readings, whether robots are blocked, we decided to utilize a simpler approach: If the
conflict cannot be solved, the priority scheme is randomly shuffled by the robot with
the highest priority, and communicated between robots that are involved in the conflict.
Subsequently, the robots try to execute the new sequence. Note that this method, as any
other method at this stage, is unable to solve a deadlock, which is a situation in which

6.2. Coordinated local exploration 113
any sequence of movements would not solve the conflict.
6.2.2 Local exploration
The fundamental problem of multi-robot exploration is how to select targets for the path
planner that minimize the overlapping of explored areas. This involves, first, choosing
a set of target locations F = { f j }, second, computing a utility value u ( f j
)
for each target
location f j ∈ F, and third, selecting the best target based on the utility value, for which
the path planner can find a trajectory.
The set of targets F is identified by extracting frontier cells F [Yamauchi, 1997] from
the occupancy grid. A frontier cell is defined as a cell that has not been explored, i.e. that
has not been observed by the laser range finder, and also neighbors an observed cell that
is unoccupied, i.e. has a occupancy probability below a certain threshold. Unexplored
cells are those with prior probability, e.g. 0.5, which is initially set for all cells. Finally,
the set is ordered based on the following utility calculation:
U ( ) ( )
f j = −γ1 F a (|φ − θ r |) − γ 2 F v l f j , (6.2)
( )
xr − x
with φ = tan −1 f j
, (6.3)
y r − y f j
where φ denotes the angle between robot location l r and frontier cell location l f j
, θ r
denotes the orientation of the robot, and γ 1 and γ 2 are two parameters which control
the trade-off between direction persistence and exploration. The function F a denotes
a fuzzy function yielding values in the interval [0, 1], the smaller the angle, the higher
the return value of the function. F v is a function returning a value the higher the more
locations in the vicinity of l f j
have been already explored. As will be shown, the information
about visited locations is read from RFIDs r ∈ LRS , which are within range of
the local occupancy grid.
The angle factor F a can be thought as an inertial term, which prevents the robot from
changing direction too often (which would result in an inefficient behavior). If the robot
has full memory of his perceptions (i.e. a global occupancy grid), the angle factor would
be enough to allow a single robot to explore the area. However, due to the limitation
of the occupancy grid, knowledge on previously visited areas gets lost, leading to inefficient
exploration. This problem is solved by memorizing explored regions in the
memory of RFIDs. If robots are within writing distance of RFIDs, they write their trajectory
into the tags memory. More specifically, they memorize visited poses p from
their trajectory (discretized at a lower resolution compared to the occupancy grid). The
influence radius, e.g. the maximal distance in which poses are added, depends mainly
on the memory capacity of the RFID tag and can be adjusted accordingly. In our implementation,
poses were added within a radius of 4 meters. Note that poses are converted
from the robot’s coordinate frame to an RFID tag relative coordinate frame in order to

114 Chapter 6. RFID Technology-based Multi-Robot Exploration
be independent of the individual coordinate frame of each robot. This has the additional
advantage that, given the resolution and maximal relative distance, their location can be
efficiently encoded into memory.
Moreover, a value count (p) [Svennebring and Koenig, 2004] is associated with each
pose p = ( x p , y p
)
in the memory of the RFID and is incremented by the robots every
time the pose is added. These poses are then used to compute F v :
F v
(
l f j
)
=
∑
r∈LRS
∑
p∈P r
count (p)
d ( l f j
, p ) (6.4)
where P r is the set of poses associated with the RFID r, and d (.) denoted the Euclidean
distance. It is worth noticing that robots writing and reading from RFIDs not only
maintain memory of their own past, but also of the other robots implementing thereby
a form of indirect communication and coordination of exploration. Thus, both multirobot
navigation and exploration do not require direct communication. This feature is
very useful in all those scenarios (e.g. disaster scenarios) where wireless communication
may be limited or unavailable. The most important feature of the approach, as presented
up to now, is that the computation costs do not increase with the number of robots.
Thus, in principle, there is no limit, other than the physical one, to the number of robots
constituting a team.
6.3 Multi-robot topological maps
After the occurrence of a real disaster it is crucial for first responders to locate and
rescue victims as fast as possible. In this context, the main task of a robot team is to
explore and map the destroyed area, and more importantly, to mark locations of victims
within the generated map. As has been already shown in Chapter 4, RFID tags can
be used for building a topological representation of the environment. The advantage
of building topological RFID maps is that this corresponding infrastructure can also be
utilized by human first responders after the robots have explored the terrain. Instead
of having to localize themselves within a metric map, which can be rather difficult if
the environment is unstructured, they can be equipped with wearable devices, such as
RFID reader and PDA, for following a path of RFID tags to task-relevant places, such
as nearby victim locations, unexplored regions, or the exit of a building.
During execution of the local exploration method, described in Section 6.2, robots
successively distribute RFIDs in the environment, which basically represent navigation
points that are connected with each other via routes traveled by the robots. This connectivity
network is taken as a basis for building a topological map consisting of vertices
that are RFID tags, starting location and victim locations, and edges that are connections
between them (see Figure 6.1 for an example). Each vertex stores the direction

6.3. Multi-robot topological maps 115
and distance to all other vertexes to which an edge exists. This graph can be used to plan
the shortest path to the next victim location and back to the exit of the building, e.g. by
utilizing the Dijkstra [Dijkstra, 1959] algorithm or A* [Russell and Norvig, 2003] planning.
In Chapter 8 a wearable device for first responders and methods for computing
optimal routes will be described in more detail.
Figure 6.1: Topological map of the disaster area generated by a team of robots in US-
ARSim environment: rectangles denote robot start locations, diamonds detected victims,
and circles RFID locations.
During the distribution of RFID tags, robots measure between two tags i and j the
relative displacement d i j = (∆x i j , ∆y i j ) T with covariance matrix Σ i j if they have been
passed. The relative displacement between two tags is estimated by a Kalman filter,
which integrates pose estimates from the robot’s wheel odometry, and an Inertia Measurement
Unit (IMU). It is assumed that the yaw angle of the IMU is aligned to magnetic
north, i.e. that IMU measurements are supported by a compass. If the robot passes a tag
or victim location, the Kalman Filter is reset in order to estimate the relative distance
to the subsequent tag on the robot’s trajectory. Accordingly, each robot generates a
graph with multiple measurements between RFID nodes, which can be exchanged with
other robots if they are within communication range. Due to the unique identification
of RFID tags, these graphs can be merged easily to a single graph consisting of the
unification of all vertices and edges, where the overall displacement D i j between two
vertices i and j can be computed from the weighted average of collected measurements
from all robots:
D i j = 1 ∑ ∑
d ki
C
j
Σ −1
k i j
, (6.5)
where
k
k
k i j
∑ ∑
C = Σ −1
k i j
(6.6)
and k i j indicates the measurement between node i and j collected by robot k. Note if
k i j

116 Chapter 6. RFID Technology-based Multi-Robot Exploration
there does not exist a measurement between two nodes, the elements of the corresponding
covariance matrix are set to zero. The method described above was applied for
building multi-robot maps during the RoboCup Rescue Virtual Competition. The generated
map can be further improved by applying RFID-SLAM, as described in Chapter 4.
6.4 Global exploration monitoring
Due to the lack of lookahead of local exploration, robots may stay too long in local
minima, resulting in redundant coverage of already explored areas. For example, this
might be the case if the utility function U ( )
l f j builds a plateau over a larger region,
i.e. if the count values of poses in a larger region are nearly the same, and unexplored
passages are only reachable with high costs from the angle factor F a .
In order to avoid such a phenomenon, a novel monitoring approach has been developed,
which periodically restarts the local exploration in more convenient locations.
This method requires direct communication and a computational overhead, which
grows with the number of agents. However, it improves the exploration ability of the
robots significantly and it is robust to failures. In fact, if the communication links
fail or the monitoring process itself fails, the robots will continue the local exploration
as previously described. The remainder of this section is structured as follows. In Section
6.4.1 the problem is formalized. Then, in Section 6.4.2 three algorithms for solving
this problem are introduced, and finally, in Section 6.4.3 the monitoring agent a central
unit that computes and monitors all global actions of the agents is introduced.
6.4.1 Problem modeling
Basically, the problem is to find optimal target locations for all robots, and a multi-robot
plan for reaching them. As a basis for solving this problem, a graph G = (V, E), where
V is the set of landmark locations, e.g. RFIDs, and E the set of connections between
them. The graph is extracted from the topological map introduced in Section 6.3. Each
node consists of a unique identifier of the RFID and its estimated position. Moreover,
a set of frontier nodes U ⊂ V and a set of current robot RFID positions S L ⊂ V is
defined. In general, |U| > |R|, where R is the set of available robots. A robot path (i.e
plan) is defined as a set of pairs composed by a node v ∈ V and a time-step t:
Definition 6.4.1 A single-robot plan is a set P = {
T = {0, . . . , |P| − 1}. P must satisfy the following properties:
a) ∀v i , v j , k ∈ P∧ ∈ P ⇒ (v i , v j ) ∈ E,
b) ∈ P ⇒ v = sl i ∈ S L
c) ∈ P ⇒ v ∈ U
| v ∈ V ∧ t ∈ T}, where

6.4. Global exploration monitoring 117
where property a) states that each edge of the plan must correspond to an edge of the
graph G. Properties b) and c) enforce that the first and the last node of a plan must
be the location of a robot and a goal node, respectively. For example, the singlerobot
plan going from RFID R1 to RFID G1, depicted in Figure 6.2, is represented
as P 1 = (< R1, 0 >, < N1, 1 >, < N2, 2 >, < G1, 3 >). The previous definition implies
1.4
G2
1
R1
0.8
N1 N2 G1
0.8 0.8
1.2
1
R2
Figure 6.2: A simple graph showing a plan from R1 to G1 (bold edges).
that passing any two nodes, which are connected by an edge in the graph G, takes approximately
the same amount of time. Recall that nodes represent RFIDs which are
deployed approximately at the same distance one from the other, and edges represent
shortest connection between them. Thus, the difference of time required for traveling
between any two connected RFIDs is negligibly small, if robots drive at the same speed.
Definition 6.4.2 A multi-robot plan P is a n-tuple of single-robot plans (P 1 , . . . , P n )
such that:
a) the plan with index i belongs to robot i,
b) ∀i, j ∈ R ∈ P i ∧ ∈ P j ⇒ v ′ v ′′
c) ∀i, j ∈ R ∈ P i ∧ ∈ P j ⇒ v ′ v ′′
Thus, a multi-robot plan is a collection of single-robot plans for each robot such that
they all have different goals and different starting positions. A distinguishing feature of
multi-robot plans with respect to single-agent ones is interaction. In fact, single-robot
plans can interfere with each other leading to inefficiencies or even failures:
Definition 6.4.3 Two single-robot plans P i and P j of a multi-robot plan P are said to
be in conflict if P i ∩ P j ∅. The set of states C Pi = ⋃ i j P i ∩ P j are the conflicting states
for P i .

118 Chapter 6. RFID Technology-based Multi-Robot Exploration
Moreover, deadlocks can occur in the system. In this setting a deadlock can arise if
there is a circular wait or if a robot is willing to move to an already achieved goal of
another robot.
Definition 6.4.4 A multi-robot plan is said to have a deadlock if there is a circular
wait or an infinite wait. The wait set for a multi-robot plan P is defined as: WS P =
{(i, j, t) | i, j ∈ R ∧ t ∈ T ∧ P i (t + 1) = P j (t)}. A circular wait for a multi-robot
plan P occurs if: ∃ a sequence of k distinct robots (r 1 , . . . , r k ) and a time t such that
(r i , r i+1 , t) ∈ WS P for i = 0 . . . k − 1 and (r k , r 1 , t) ∈ WS P . An infinite wait for a multirobot
plan P occurs if ∃t 1 , t 2 ∈ T ∃i, j ∈ R | P i (t 1 ) = P j (t 2 ) ∧ t 2 = |P j | − 1 ∧ t 1 ≥ t 2 .
Consequently, the cost measure c(.) for a multi-robot plan P is defined as follows:
⎧
⎪⎨ ∞
if deadlock
c(P) = ⎪⎩ max cost(P, i) else (6.7)
i∈R
where cost(P, i) is the cost of executing i’s part of the multi-robot plan P. We assume
that the agents execute the plans in parallel, thus the score of the multi-robot plan is the
maximum among the single-robot ones. Let P j (t) be a function that returns the RFID
node of a plan P j at a time index t, and d(.) the Euclidean distance between two RFIDs.
Then, cost(P, i) can be computed from the sum of the Euclidean distances between the
RFIDs of the plan plus the conflicts cost:
where
and
cost(P, i) =
|P∑
i |−2
t=0
∑
con f l(P, i) =
d(P i (t), P i (t + 1)) + con f l(P, i) (6.8)
ji
∑
∈P i ∩P j
wait(P j , t), (6.9)
wait(P j , t) = d(P j (t − 1), P j (t)) + d(P j (t), P j (t + 1)), (6.10)
where the wait cost wait(P j , t) reflects the time necessary for robot j to move away from
the conflict node. By Equation 6.9 costs for waiting are added if at least two robots share
the same RFID node at the same time. This is a worst case assumption since conflicts in
the final multi-robot plan are solved by the local coordination mechanism which forces
robots only to wait if there are other robots with higher priority. We abstract this feature
from our model since the priority ordering is periodically randomized in order to solve
existing dead-locks, making it impossible to predict whether a robot will have to wait
or not. Finally, the task assignment and path planning problem can be formulated as an
optimization problem of finding a plan P ∗ that minimizes the cost function c(P).

6.4. Global exploration monitoring 119
6.4.2 Global task assignment and path planning
We experimented with three different techniques in order to solve the task assignment
and path planning problem. The first two approaches are inspired by Burgard et al. [Burgard
et al., 2005]. The third approach can be seen as an extension of Bennewitz et
al. [Bennewitz et al., 2001]. All of the previously cited approaches rely on a grid based
representation while our approach is graph-based. The experimental results show that
the third approach outperforms the first and the second, and is actually the one we
adopted in the implementation of the full system. For this reason, only a brief overview
of the first two approaches is given, but a more detailed description of the third.
All the approaches have a common pre-calculation. We compute the pairwise shortest
paths [Dijkstra, 1959] for each node in U. This is a fast computation (i.e. O((|E| +
|V|log(|V|))|U|) ) which speeds up the plan generation processes presented in the following.
Greedy Approach
Given the information produced by the Dijkstra algorithm and an empty multi-robot
plan, we identify the robot r best ∈ R which has the shortest path to reach a goal g best ∈ U.
The path computed by the Dijkstra algorithm from r best to g best , with its time values is
added to the multi-robot plan. We then update the sets R = R−{r best } and U = U−{g best }.
The process is iterated until R = ∅ (see [Burgard et al., 2005] for more details).
Assignment Approach
Task assignment is a common approach in multi-robot systems to coordinate tasks between
the agents. Here we utilize a genetic algorithm permuting over possible goal
assignments to robots and use the plans computed by the Dijkstra algorithm. We then
use the previously defined cost function as the fitness function (see [Bennewitz et al.,
2001] for more details).
Sequential Approach
The last presented approach is based on sequential planning. We use, similar to the
assignment approach, a genetic algorithm to permute possible orderings of agents O =
o 1 , . . . , o n . We then plan the ordering and use the previously defined cost function as
the fitness function.
The sequential planning is based on A* [Russell and Norvig, 2003] and is done individually,
following the given sequence, for each agent in order to achieve the most
convenient of the available goals U. Every time when an agent o i plans, the selected
goal is removed from U and the computed plan added to the set of known plans P. The
planning tries to avoid conflicts with the set of known plans P by searching through

120 Chapter 6. RFID Technology-based Multi-Robot Exploration
time/space, where the state space S is defined as S = V × T. This huge state space can
be greatly simplified since we are only interested in the time of the conflicting states
C P j
(Definition 6.4.3). From the planning point of view the information relative to the
time of non-conflicting states is irrelevant and thus all these states can be grouped by
time using the special symbol none. The resulting set of non-conflicting states NC is
defined as NC = { |v ∈ V} and has the cardinality of V (i.e. |NC| = |V|). Thus,
the reduced search space is S T = C P j
∪ NC.
During the search, the nodes are expanded in the following way: We look for the
neighboring nodes of the current one given the set E of edges in G. For each of them
we check if there is a conflict. If this is the case, we return the corresponding node from
C P j
, otherwise the one from NC.
In order to implement A* we have to provide a cost function g and a heuristic function
h defined over S T. We define the cost function g(s) for agent i as the single-robot
plan cost function cost(P, i). The multi-robot plan P will consist of the plans already
computed with the path found up to s. Obviously the agent o j will be able to detect
conflicts at planning time only for those agents o i with i < j for which a plan has
already been produced. Finally, the heuristic h (i.e. Dijkstra heuristic) is defined as
follows:
h(< s, t >) = min d di j(s, g) (6.11)
g∈G
where d di j (s, g) is the distance from s to g pre-computed by the Dijkstra algorithm,
which is independent of time t. It is important to notice that A* will find the optimal
solution since the heuristic is admissible, i.e. it underestimates the true costs, for a
single robot plan given the subset of already computed paths and ignoring the others
(for which no plan was found yet).
For example, let us consider the simple weighted graph depicted in Figure 6.2. R1
and R2 are respectively the location of two robots r1 and r2. G1 and G2 are the goals.
In this example, the sequence has been selected and r1 has already produced
the following plan: (< R1, 0 >, < N1, 1 >, < N2, 2 >, < G1, 3 >). Now r2 has to plan.
The only remaining goal is G2 since G1 has already been selected by r1. At first,
according to the topology and the already defined plan for r1, from the
nodes and < N1, 1> can be reached. In fact, if we simulate the plan of
r1, moving to R1 will entail no conflict and would have just the cost of traveling the
distance; thus g() = 1.2. In the other case, moving to R2 at time 1, will
conflict with r1 who is moving there at the same time. In this case, our model will tell
us that we have to wait for r1 to first reach N1 (with a cost of 0.8) and then leave it
(with a cost of 0.8). Then, we would be able to reach N1 with a cost of 1. Thus the total
cost of reaching N1 at time 1 will be g() = 2.6 (i.e. 0.8 + 0.8 + 1).
Moreover, the heuristic values for these states (obtained by pre-computing the Dijkstra
algorithm) are: h() = 1.4 and h(< N1, 1>) = 1. Thus, according to
the well-known formula f (n) = g(n) + h(n), will be selected. Similarly,

6.4. Global exploration monitoring 121
nodes and will be expanded next and the planning process will
continue until the plan (, , ) is found. Notice that,
is different from since r2 will already have moved away from N1 at time-step
2.
6.4.3 Monitoring agent
The monitoring agent MA constructs the map represented as the graph G online and
identifies the frontier RFIDs, which are RFIDs at incompletely explored areas. Moreover,
MA will monitor the local exploration and possibly identify, with one of the previously
described techniques, a multi-robot plan in order to move robots to locations
offering better performance expectations for the local exploration.
At execution time the robots send to MA their RFID locations (i.e. the nearest RFID
they can perceive). Every time a robot changes its RFID position from r i to r i+i , MA
updates the set S L of current robot locations and updates the graph as follows: E =
E ∪ {(r i , r i+1 )} and V = V ∪ {r i } ∪ {r i+1 }.
The monitoring process collects information continuously regarding the unexplored
area in the vicinity of the RFIDs based on the local occupancy grid to identify the
frontier RFIDs U. Roughly, the robot knows how many RFIDs, given the defined deployment
density, should be placed per square meter and which is the density they
perceive. Therefore, they are able to compute an estimate on how much the area has
been explored in the proximity of the RFIDs they visit.
MA periodically evaluates the position of the robots on the graph and their distance
from the frontier nodes U. If this value exceeds a given threshold, it stops the robots
and computes a new multi-robot plan. Once a valid plan has been produced, MA starts
to drive the robots by assigning to each of them the next RFID prescribed by their
plans. Robots plan between RFIDs by using A* on the occupancy grid and the teammate
avoidance, as described in Section 6.2.1, but with RFID locations as goals rather
than frontier cells. If the plan execution fails due to unobservable RFIDs or obstructed
connections, a failure message is sent to MA. MA will consequently remove the node
and its edges from the graph G and re-plan. When the target RFID is reached, a task accomplished
message is sent to MA, which will assign another task or send a global plan
termination message. In the latter case, the robots will start again with local exploration.
Further, during the multi-robot plan execution, the planner monitors exploration for
detecting unforeseen situations. For example, if a robot does not send an accomplished
task message or an RFID position for a long time, it is considered lost and removed
from the robot list. During multi-plan execution deadlocks might occur. This is due to
the fact that the current system does not allow to predict the exact order in which tasks
will be accomplished. This issue is currently solved by re-planing the assignment if a
deadlock occurs, or by restarting local exploration if no better plans can be found.

122 Chapter 6. RFID Technology-based Multi-Robot Exploration
6.5 Experiments
We utilized a simulated model of the Zerg robot, that captures the same physical properties
as the real one, e.g. a four wheel drive, an RFID tag release device, an RFID
antenna, Inertial Measurement Unit (IMU), and LRF. The sensors of the model are simulated
with the same parameters as the real sensors, expect the real RFID reading and
writing range. Without loss of generality, these ranges have been extended to two meters,
since this parameter mainly depends on the size of the transmitter’s antenna, which
can also be replaced.
6.5.1 Evaluation of the local approach
The local approach has been tested intensively with the USARSim simulator ( [Carpin
et al., 2006a, Carpin et al., 2006b, Balakirsky et al., 2006]) for various environments
generated by the National Institute of Standards and Technology (NIST). They provide
both indoor and outdoor scenarios of the size bigger than 1000 m 2 , reconstructing the
situation after a disaster. Since USARSim allows for the simulated deployment of heterogeneous
robot types in a wide range of different scenarios, it offers an ideal performance
metric for comparing multi-robot systems. Furthermore, the effort of evaluating
large robot teams is greatly reduced. The local approach was applied in USARSim during
the RoboCup’06 Virtual Robots Competition, where our team won the first prize ∗ .
During this competition, virtual teams of autonomous or tele-operated robots had to
find victims within 20 minutes while exploring an unknown environment. The simulation
system was capable of simulating up to 12 robots at the same time. Most of
the other teams applied frontier cell-based exploration on global occupancy grids. In
particular: selfish exploration and map merging [Nevatia et al., 2006] (IUB), map merging
and local POMDP planning [Pfingsthorn et al., 2006] (UVA), operator-based frontier
selection and task assignment (SPQR), and tele-operation (STEEL) and (GROK).
The competition was held in three preliminary rounds, two semifinals and two finals.
In the following the results in terms of exploration and victim discovery of our team
RRFreiburg compared to the results of other teams at the competition are presented.
Exploration was evaluated based on the totally explored area of each team. These
values were automatically computed from the logs of the server hosting the simulator.
Table 6.1 gives an overview of the number of deployed robots, and the area explored by
each team. It also provides additional information such as the ratio between the totally
explored area and the summed distance traveled by all robots (Area/Length). Furthermore,
the average area explored by a single robot of each team is shown (Area/#robots).
The result shows clearly that our team was able to deploy the largest robot team, while
exploring an area bigger than any other team. Due to the modest computational re-
∗ Virtual Rescue Robots Freiburg: www.informatik.uni-freiburg.de/∼rescue/virtual

6.5. Experiments 123
RRFreiburg GROK IUB SPQR STEEL UvA
Preliminary 1 # Robots 12 1 6 4 6 1
Area [m 2 ] 902 31 70 197 353 46
Area/Length 0,65 1,33 0,49 0,89 0,98 0,82
Area/#robots 75,17 31 11,67 49,25 58,83 46
Preliminary 2 # Robots 12 1 4 4 6 8
Area [m 2 ] 550 61 105 191 174 104
Area/Length 0,35 0,92 0,7 0,9 0,83 0,56
Area/#robots 45,83 61 26,25 47,75 29 13
Preliminary 3 # Robots 10 1 5 7 6 7
Area [m 2 ] 310 59 164 44 124 120
Area/Length 0,38 1,02 0,64 0,88 0,53 0,41
Area/#robots 31 59 32,8 6,29 20,67 17,14
Semifinal 1 # Robots 8 1 6 4 6 6
Area [m 2 ] 579 27 227 96 134 262
Area/Length 0,23 0,68 0,88 0,67 0,7 0,72
Area/#robots 72,38 27 37,83 24 22,33 43,67
Semifinal 2 # Robots 8 1 6 5 6 7
Area [m 2 ] 1276 82 139 123 139 286
Area/Length 0,51 1,03 0,91 0,99 0,51 0,71
Area/#robots 159,5 82 23,17 24,6 23,17 40,86
Final 1 # Robots 8 - 8 - - -
Area [m 2 ] 1203 - 210 - - -
Area/Length 0,47 - 0,93 - - -
Area/#robots 150,38 - 26,25 - - -
Final 2 # Robots 8 - 6 - - -
Area [m 2 ] 350 - 136 - - -
Area/Length 0,2 - 0,53 - - -
Area/#robots 43,75 - 22,67 - - -
Table 6.1: Exploration Results from RoboCup ’06
sources needed by the local approach, we were able to run 12 robots on a single Pentium4,
3 GHz.
Victim points were awarded for both locating victims and providing extra information.
Table 6.2 summarizes the victim information reported by each team. In particular,
the table shows the number of victims identified, the bonus for reporting the status (i.e
10 points for each victim status) extra information such as pictures (up to 20 points for
each report), and accurate localization of the victims (up to 20 points).
Figure 6.3 (a-b) depicts the joint trajectory of each team generated during the semifinal
and final and (c-d) shows the single trajectory of each robot of our team on the
same map, respectively. The efficiency of the RFID-based coordination is documented
by the differently colored trajectories of each single robot.

124 Chapter 6. RFID Technology-based Multi-Robot Exploration
RRFreiburg GROK IUB SPQR STEEL UvA
Preliminary 1 reported victims 5 2 1 1 2 0
status bonus 20 0 10 0 20 0
victim bonus - - - - - -
Preliminary 2 reported victims 2 1 1 3 2 1
status bonus 20 0 10 30 20 0
victim bonus - - 15 - 30 -
Preliminary 3 reported victims 5 3 3 1 5 3
status bonus 50 0 0 0 50 0
victim bonus - 45 30 - 60 -
Semifinal 1 reported victims 8 4 7 6 7 6
status bonus 70 20 30 50 60 30
victim bonus - 60 105 45 90 -
localization bonus 80 20 70 48 70 60
Semifinal 2 reported victims 16 4 5 9 14 2
status bonus 80 0 40 60 140 0
victim bonus - 45 75 75 0 -
localization bonus 160 40 50 90 140 20
Final 1 reported victims 10 - 3 - - -
status bonus 60 - 30 - - -
victim bonus - - 60 - - -
localization bonus 100 - 30 - - -
Final 2 reported victims 10 - 5 - - -
status bonus 80 - 10 - - -
victim bonus - - 30 - - -
localization bonus 100 - 50 - - -
Table 6.2: Victim scoring results from RoboCup ’06
6.5.2 Evaluation of the global approach
Efficiency in terms of conflict detection and joint path length optimization was evaluated
on the basis of both artificially generated, and robot team generated RFID graphs.
The artificially generated graphs, consisting of approximately 100 nodes, are weakly
connected in order to increase the difficulty of the planning problem, whereas the graph
generated by the robots, consisting of approximately 600 nodes, represents a structure
naturally arising from an office-like environment.
Figure 6.4 depicts the result from evaluating greedy assignment, genetic optimized
assignment, and sequence optimization on these graphs. Each method was applied with
a fixed number of randomized goals and starting positions 10 times. We experimented
with different sizes of robot teams ranging from 2 to 20. The abrupt ending of the
curves indicate the size of the agent team, at which no more solutions could be found,
i.e. the scoring function returned infinity. Note that for all the experiments, the genetic
algorithm was constrained to compute for no more than one second.
The result makes it clear that sequence optimization helps to decrease both the over-

6.5. Experiments 125
(a)
(b)
(c)
(d)
Figure 6.3: Exploration scoring trajectories recorded during the finals: (a,b) Comparison
between our approach (red line) and all other teams. (c,d) Coordinated exploration
of our robots with each robot represent by a different color.
all path costs and the number of conflicts between single robot plans. Moreover, the
method yields solutions with nearly no conflicts on the graph dynamically generated by
the robot team (see Figure 6.4 (c)). In order to compare the global and local approach
in terms of the explored area, we conducted two experiments on a large map, for 40
minutes each. Due to the global approach, the robots were able to explore 2093 m 2 of
the map, in contrast to the team executing the local approach, exploring only 1381 m 2 of
the area. The trajectories in Figure 6.5 indicate that this was mainly because the robots
running the local approach were not able to overcome the local minima in the long hall.
With the global approach, the robots discovered the passage leading to the large area
beneath the hall.

126 Chapter 6. RFID Technology-based Multi-Robot Exploration
18
18
16
14
Greedy
Sequence opt.
Assignment opt.
16
14
# Conflicts
12
10
8
6
# Conflicts
12
10
8
6
Greedy
Sequence opt.
Assignment opt.
4
4
2
2
0
2 4 6 8 10 12 14 16 18 20
# Robots
(a)
0
2 4 6 8 10 12 14 16 18 20
# Robots
(b)
# Conflicts
20
18
16
14
12
10
8
6
4
2
Greedy
Sequence opt.
Assignment opt.
0
2 4 6 8 10 12 14 16 18 20
# Robots
(c)
Costs
900
800
700
600
500
400
300
200
Greedy
Sequence opt.
Assignment opt.
100
2 4 6 8 10 12 14 16 18 20
# Robots
(d)
800
700
600
Greedy
Sequence opt.
Assignment opt.
1600
1400
1200
Greedy
Sequence opt.
Assignment opt.
Costs
500
400
Costs
1000
800
300
600
200
400
100
2 4 6 8 10 12 14 16 18 20
# Robots
(e)
200
2 4 6 8 10 12 14 16 18 20
(f)
# Robots
Figure 6.4: Comparing the number of conflicts (a-c) and travel costs (d-f) of the three
approaches on different RFID graphs: (a,d) narrow office-like environment, (b,e) narrow
outdoor area, (c,f) graph generated from RFIDs deployed by the robots on a USARSim
map.

6.6. Related work 127
(a)
(b)
Figure 6.5: Comparing the locally and globally coordinated exploration. During local
exploration (a) robots get stuck in a local minima. The global approach (b) allows the
robots to leave the local minima and to explore a larger area
6.6 Related work
There has been a wide variety of approaches to robotic exploration, whereas most of
them neither consider the optimization of multi-robot plans, nor do they directly address
the problem of reliable map sharing between robots. Furthermore, they require for map
exchange a minimal amount of communication bandwidth.
Zlot et al. proposed an approach for coordinating multi-robot exploration guided by a
market economy [Zlot et al., 2002]. Their method trades exploration tasks using singleitem
first-price sealed-bid auctions between the robots. Although the system is partially
functional with low bandwidth communication, it is unclear by which extend exploration
efficiency decreases if communication is significantly perturbed. Methods for
local exploration have already been successfully applied in the past [Balch and Arkin,
1994, Svennebring and Koenig, 2004]. It basically has been shown that multi-robot
terrain coverage is feasible without robot localization and communication. In their
scenario, robots coordinate their tracks by leaving physical markings in the environment.
However, it is unclear whether these markings can be reliably detected by the
robots. Furthermore, physically markings are not desirable in every scenario. Burgard
et al. [Burgard et al., 2005] contributed a method for greedy task assignment based on
grid mapping and frontier cell exploration [Yamauchi, 1997]. Their approach trades off
the information gain if reaching frontier cells with the cost of moving to them. They
do not consider conflicts between single robot plans, and require robots to start their
mission close to each other with knowledge about their initial displacement. Also the
idea of compressing maps for reducing the amount of data during communication has
been considered [Meier et al., 2005]. In this work, the approximation of grid maps

128 Chapter 6. RFID Technology-based Multi-Robot Exploration
by polygons has been evaluated. The work by Bennewitz and colleagues [Bennewitz
et al., 2001] focuses on the optimization of plans taken by multiple robots at the same
time. They select priority schemes by a hill-climbing method that decides in which
order robots plan to their targets [Erdmann and Lozano-Perez, 1987]. Plans are generated
in the configuration time space by applying A* search on grid maps. Farinelli et
al. considered a scenario where tasks to be accomplished are perceived by the robots
during mission execution. [Farinelli et al., 2006]. The paper presents an asynchronous
distributed mechanism based on Token Passing for allocating tasks in a team of robots.
The coordinated movement of a set of vehicles has also been addressed in other
domains, such as in the context of operational traffic control [Hatzack and Nebel, 2001],
and the cleaning task problem [Jaeger and Nebel, 2001].
6.7 Conclusion
In this chapter a novel method for coordinated exploration of large robot teams has
been contributed. The approach, which is based on RFID technology for indirect communication,
is composed of two parts. First, distributed local search, with the notable
properties of not requiring direct communication and scaling with the number of agents.
Second, global task assignment and multi-robot path planning for monitoring the local
exploration, and restarting it in better locations.
The usage of RFIDs allows to build a significantly smaller representation of the environment
compared to grid based approaches [Nevatia et al., 2006, Pfingsthorn et al.,
2006, Burgard et al., 2005, Bennewitz et al., 2001]. The experimental results from
RoboCup show that RFID-based coordination scales with large robot teams exploring
large environments.
Moreover, we extensively experimented with three approaches for solving the task
assignment and planning problem. The first two are genetic and greedy task assignment
techniques [Burgard et al., 2005], whereas the third is a variant of sequential
planning [Bennewitz et al., 2001]. It has been shown that genetic optimization, particularly
sequence optimization, outperforms the greedy assignment in terms of conflict
reduction and path length. With increasing number of robots, greedy assignment fails to
provide executable solutions, whereas sequence optimization is able to solve problems
with up to 20 robots. Finally, qualitative experiments with the full system have shown
that the global method might lead to the exploration of areas that are more than double
in size as achieved by the local approach.
In Chapter 4 experiments have been conducted that demonstrated the capability of
the real-robot platform Zerg to autonomously deploy and detect RFID tags in a cellar
environment. Since the simulated robots are running the same software modules as the
real robots, we are confident that the proposed exploration system is transferable to the
real-robot platform.

6.7. Conclusion 129
There are several issues that have to be considered by future work. First, the approach
has to be evaluated on a real robot team exploring a larger area. Second, the plan
execution phase of the approach can be further improved. It is considered to represent
the generated multi-robot plans by Petri nets, allowing to verify online plans generated
from task assignment and path planning [Ziparo and Iocchi, 2006]. Specifically multirobot
plans are represented as a discrete event system with RFIDs as resources that may
be owned by one robot at the time. Given this representation, plan executions can be
simulated for computing a more precise plan evaluation. On the one hand, this facilities
the optimization of robot waiting times if they are involved within conflicts. On the
other hand, it enables deadlock detection during the planning phase. Finally, the model
allows to compute deadlock-free execution policies that can replace the randomized
priority ordering.

7 Victim Detection
7.1 Introduction
Robots deployed for exploration and mapping of an area after a disaster are required
to perform subtasks autonomously as much as possible. One primary goal is to deploy,
under the surveillance of a human operator, a team of robots for coordinated victim
search. This requires robots to perform subtasks, such as victim detection, partially or
even fully autonomously.
The National Institute of Standards and Technology (NIST) develops test arenas for
the simulation of situations after a disaster [Jacoff et al., 2001]. In this real-time scenario,
robots have to explore an unknown area autonomously within 20 minutes and to
detect victims therein. There might be “faked” victim evidence, such as printed images
of human faces, non-human motion, and heat sources that do not correspond to victims.
The heat blanket in the fourth row of Figure 7.4, for example, would be wrongly
reported as victim by most heat-seeking robots. Note that this example is particularly
difficult due to the large size of the thermo signature, as well as the closely located
evidence given by the skin-like color of the blanket, the face, and motion.
Due to the real-time constraint in rescue-like applications, only fast computable
techniques are admissible. During RoboCup’05 and RoboCup’06 the RescueRobots
Freiburg team successfully applied simple but fast classifiers for victim detection, such
as color thresholding, motion detection, and shape detection on images taken by an infrared
and color camera, respectively. However, the detection rate of these classifiers
turns out to be moderate, since they are typically tuned for specific objects found in
the environment. Hence, in environments containing many diverse objects, they tend
to produce a large number of evidence frames, from which in the worst case most are
false-positives, i.e objects that are wrongly recognized as victims. One solution to this
problem is to combine local evidences, i.e. evidences that are close to each other in
the real world, and to reason on their true class label with respect to their neighborhood
relations. Markov Random Fields (MRFs) provide a probabilistic framework for
representing such local dependencies. However, inference in MRFs is computationally
expensive, and hence not generally applicable in real-time.
In this chapter a novel approach for the genetic optimization of the building process of
MRF models is contributed. The genetic algorithm determines offline relevant neighborhood
relations, for example the relevance of the relation between evidence types
131

132 Chapter 7. Victim Detection
heat and motion, with respect to the data. These pre-selected types are then utilized
for generating MRF models during runtime. First, the vertices of the MRF graph are
constructed from the output of the weak classifiers. Second, edges between these nodes
are added if the specific type of nodes can be connected by an edge type that has been
selected during the optimization procedure.
Experiments carried out on test data generated in environments of the NIST benchmark
indicate that compared to the utilized Support Vector Machine (SVM) based classifier,
the optimized MRF models reduce the false-positive rate. Furthermore, the optimized
models turned out to be up to five times faster than the non-optimized ones at
nearly the same detection rate.
The remainder of this chapter is structured as follows. In Section 7.2 the underlying
vision system is introduced, Section 7.3.1 explains the MRF model, and in Section 7.3.2
the genetic model selection approach is introduced. In Section 7.5 related work is discussed.
Finally, results from experiments are presented in Section 7.4, and the chapter
is concluded in Section 7.6.
7.2 Vision data processing
The utilized vision system is part of the rescue robot Zerg, shown in Figure 7.1 (a),
which is equipped with a Hokuyo URG-X004 Laser Range Finder (LRF), a ThermalEye
Infra-Red (IR) camera, and a Sony DFW-V500 color camera. The LRF is capable of
measuring distances up to 4000 mm within a field of view (FOV) of 240 ◦ and the FOV
of the IR and color camera are 50 ◦ and 70 ◦ , respectively. Please see Section 2.5 for
further details.
In order to combine evidence from thermo and color images, we first project their
pixels onto the 3D range scan, and second determine pixel-pixel correspondence by interpolating
from best matching yaw and pitch angles found in both projections. Before
camera images are projected onto the scan, they are linearized with respect to the intrinsic
parameters of the camera. On color cameras, these parameters are usually calibrated
from pixel to real-world correspondences generated by a test pattern, such as the printout
of a chess board [Bradski, 2000]. In case of IR camera calibration, it is necessary
to generate a test pattern that also appears on thermo images. This has been achieved
by taking images from a heat reflecting metal plate covered with quadratic isolation
patches in a chess board-like manner.
From both images three different evidence types are generated, which are color, motion,
and shape, respectively. Each evidence type is represented by a rectangular region
described by the position and size (u, v, w, h) on the image, number of pixels included,
and the real world position (x, y, z) of the center.
Color pixels are segmented by fast thresholding [Bruce, 2000] in the YUV color
space. In case of the IR camera, only the luminance (brightness) channel is used since

7.2. Vision data processing 133
(a)
(b)
(c)
(d)
Figure 7.1: The autonomous rescue robot Zerg (a) and vision data from the same scene
(b)-(d): A color image taken by the CCD camera (b), a thermo image taken by the IR
camera (c), and a 3D scan taken by the 3D scanner (d).
thermo images are represented by single values per pixel, which are proportional to the
detected temperature. Pixels within the same color class are merged into blobs by run
length encoding, and represented by rectangular regions.
Motion is detected by background subtraction of subsequent images. Let I t be an
image at time t from a sequence of images I with I 0 = Background. Then, the difference
between an image and the background can be calculated by AVG t = (1 − β) AVG t−1 +
βI t , DIFF t = AVG t −I t , where AVG is the running average over all images and β a factor
controlling the trade-off between latency and robustness. Pixels labeled as foreground
are also merged into groupings by run length encoding and are represented by a set of
rectangular regions.
Shape detection is currently limited to the detection of human faces. We use the
openCV [Bradski, 2000] implementation of the method from Viola and colleagues [Viola
and Jones, 2001], which has been further improved by Lienhart [Lienhart and

134 Chapter 7. Victim Detection
Maydt, 2002]. The method utilizes a cascade of haar-like features that are trained and
boosted from hundreds of sample images scaled to the same size. Since the classifier
was mainly trained from images with faces aligned in the vertical direction, we rotated
images for allowing the detection of faces aligned horizontally.
7.3 Genetic model optimization
7.3.1 Markov Random Fields (MRFs)
The series of images in Figure 7.4 (a) clearly show that single evidences are not sufficient
to uniquely identify victims. Therefore, it is necessary to consider neighborhood
relations in order to reduce false-positive detections. Markov Random Fields (MRFs)
provides a probabilistic framework for representing local dependencies. A MRF is
defined by an undirected graph G = (Y, E), where Y is a set of discrete variables
Y = {Y 1 , . . . , Y N }, and E is a set of edges between them. Each variable Y i ∈ {1, . . . , K}
can take on one of K possible states. Hence, G describes a joint distribution over
{1, . . . , K} N .
According to the approach of Anguelov et al. [Anguelov et al., 2005], we utilize pairwise
Markov networks, where to each node a potential φ (y i ) and to each undirected edge
E = {(i j)} (i < j) between two nodes a potential φ ( )
y i , y j is associated. Consequently,
the pairwise MRF model represents the joint distribution by:
P φ (y) = 1 Z
N∏ ∏
φ i (y i ) φ i j (y i , y j ), (7.1)
i=1 (i j)∈E
where Z denotes a normalization constant, given by Z = ∑ y ′ ∏ N
i=1 φ i(y ′ i ) ∏ (i j)∈E φ i j (y ′ i , y′ j ).
A specific assignment of values to Y is denoted by y and represented by the set { }
y k i of
K · N indicator variables, for which y k i
= I(y i = k). In order to foster the associativity of
the model, we reward instantiations that have neighboring nodes, which are labeled by
the same class. This is enforced by requiring φ i j (k, l) = λ k i j , where λk i j
> 1 for all k = l,
and φ i j (k, l) = 1 otherwise [Taskar et al., 2004]. Inference is carried out by solving the
( )
maximum a-posterior (MAP) inference problem, i.e. to find arg max y P φ y .
For our specific problem, we define the node potentials φ (y i ) by a vector of features
which indicates the quality of the node’s corresponding evidence frame. These features
are the size of the evidence frame, the number of included pixels, and the real world
distance taken from the 3D range measurement. Likewise we define the edge potentials
φ ( )
y i , y j by a vector of features that indicates the quality of neighborhood relations.
Edges are build from combinations of the evidence types introduced in Section 7.2. In
our case there exist 36 possible edge types, given the 6 different types of evidence. For

7.3. Genetic model optimization 135
example, the edge type isWarmSkin describes the combination of the features heat and
skinColor. The feature vector of an edge includes the type of the edge and the realworld
distance measurement between both nodes. Some edge features are shown in
Table 7.1.
IR-Heat IR-Motion IR-Face
IR-Heat isSameType isWarmMotion isWarmFace
IR-Motion - isSameType isMovingFace
IR-Face - - isSameType
Color-Skin - - -
Color-Motion - - -
Color-Face - - -
Color-Skin Color-Motion Color-Face
IR-Heat isWarmSkin isWarmMotion isWarmFace
IR-Motion isMovingSkin isWarmMotion isMovingFace
IR-Face isSkinFace isMovingFace isWarmFace
Color-Skin isSameType isMovingSkin isSkinFace
Color-Motion - isSameType isMovingFace
Color-Face - - isSameType
Table 7.1: Feature combinations of evidence types heat/skin, motion, and face, generated
from color images and thermo images, respectively.
The MRF graph is dynamically constructed for each image from the video stream
during runtime (see Figure 7.2 for an example). First, we generate from the image
data six sets of evidence frames, as described in Section 7.2. From these sets, six
types of MRF nodes are generated by calculating the feature vectors for each node
potential, where each node type corresponds to one type of evidence. Second, edges
between nodes are generated. Each node connects to the four closest neighbors in its
vicinity if they are within close real-world distance, which was maximally 60cm in
our implementation. Finally, for each edge a feature vector for its edge potential is
calculated.
For the sake of simplicity, we represent potentials by a log-linear combination log φ i (k) =
w k n · x i and log φ i j (k, k) = w k e · x i j , where x i denotes the node feature vector, x i j the edge
feature vector, and w k n and w k e the row vectors according to the dimension of node features
and edge features, respectively. Consequently, we can denote the MAP inference
( )
problem arg max y P φ y by:
arg max
y
N∑ K∑
∑ K∑
(w k n · x i )y k i + (w k e · x i j )y k i yk j . (7.2)
i=1 k=1
(i j)∈E k=1

136 Chapter 7. Victim Detection
Figure 7.2: MRF graph online constructed from features detected in the thermo and
color images. Note that the number of shown features has been reduced for the sake of
readability.
Equation 7.2 can be solved as a linear optimization problem by replacing the quadratic
term y k i yk j
with the variable y k i j and adding the linear constraints yk i j
≤ y k i
and y k i j
≤ y k j .
Hence, the linear programming formulation of the inference problem can be written as:
max
N∑
i=1
K∑
∑
(w k n · x i )y k i +
k=1
s.t. y k i ≥ 0, ∀i, k;
(i j)∈E k=1
K∑
(w k e · x i j )y k i j (7.3)
∑
y k i = 1,
y k i j ≤ yk i , yk i j ≤ yk j , ∀i j ∈ E, k,
which can, for example, be solved by the Simplex method. Furthermore, it is necessary
to learn the weight vectors for the node and edge potential from data, which has been
carried out by utilizing the maximum margin approach recommended by Taskar and
colleagues [Taskar et al., 2003].
k
∀i;
7.3.2 Model selection
In general, solving the MAP inference problem, as shown in Section 7.3.1, is NPhard
[Shimony, 1994]. Also in case of the considered two-class problem one notices
an increase in computation time if the number of nodes and edges, and thus size of the

7.3. Genetic model optimization 137
linear programming problem, grows.
Computation time is an important issue if applying the detection method, for example,
to live video streams taken by a camera in a Search and Rescue scenario. Furthermore,
the efficiency of specific evidence correlations (edge types in the MRF graph)
depends on the scenario where the classifier is applied. For example, heat sources might
be a stronger evidence for human bodies in an outdoor scenario as it would be in an indoor
scenario with many heat sources, such as PCs and radiators. Therefore, our goal is
to reduce the computation time of the MRF model by selecting edge types that significantly
improve the classifier with respect to the data. For example, measurements from
different sensor types are generally more significant than measurements from a single
sensor type. The two edges between the motion node and the two heat nodes in Figure
7.2 are more valuable than a single connection between heat nodes only. However,
by selecting the four closest nodes both heat nodes would be connected.
Therefore, we examined the contribution of specific edge types to the overall detection
rate. This has been carried out by learning MRF models with different sets of
activated edge types while measuring accuracy and computation time needed for inference.
Figure 7.3 summarizes the result, where each data point corresponds to a specific
combination of edge types. MRF inference needs between 2 ms and 16 ms, depending
100
80
Correct classified [%]
60
40
20
False Positives
Correct classified
False Positive [%]
0
0 2 4 6 8 10 12 14 16
Computation Time [ms]
Figure 7.3: Comparing model complexity (computation time) with the percentage of
correctly classified vision features (green), and the percentage of false positives (red).
Each data point corresponds to a MRF model with specific types of edges activated.
on the combination of activated edge features. Surprisingly, a higher amount of computation
time does not necessarily yield better classifier performance. Good classifier
performance can be achieved already at a much smaller computation time than needed

138 Chapter 7. Victim Detection
for computing models containing all types of edges if the significant edge types are
activated.
However, since the complexity of exhaustive search is in O (2 n ), and learning a single
classifier takes a comparably high amount of time, finding optimal edge types is
intractable in the general case. Therefore, we applied a genetic algorithm for selecting
the most efficient combinations of edges. The scoring function for guiding the search
has been defined by the trade-off between classifier performance and computation time:
S = U − αC (p i ) , (7.4)
where U corresponds to the utility metric, C (.) denotes a cost function reflecting the
model complexity, p i denotes the ith permutation, and α is a parameter regulating the
trade-off. Depending on the needs of the specific application, U can be computed, for
example, from the negative false-positive rate, the total number of correctly classified
evidence frames, and the percentage of the correctly classified area. Without loss of
generality, we decided to use the area-based utility metric since it enforces the detection
of body silhouettes rather than frames on their own. The series in Figure 7.4 (c) depicts
this metric for true positives by the blue cluster, which was build by the union of all true
positive evidence frames in the image.
The scoring function is utilized as fitness function for the genetic algorithm (GA).
Solutions, i.e. specific combinations of edge types, are represented for the genetic optimization
as a binary string. Each edge type is represented by a bit, and set to true or
false regarding the activation of the corresponding edge type. In order to guarantee that
good solutions are preserved within the genetic pool, the so-called elitism mechanism,
which forces the permanent existence of the best found solution in the pool, has been
used. Furthermore, we utilized a simple one-point-crossover strategy, a uniform mutation
probability of p ≈ 1/n, and a population size of 10. In order to avoid that solutions
are calculated twice, all computed solutions are memorized by their binary string in a
hash map.
7.4 Experiments
We generated more than 6000 labeled examples from video streams recorded in a NIST
arena-like environment and split them into three folds. The training data contained true
evidence which is exclusively generated from human bodies and false evidence, which
is generated from artificial sources, such as a heat blanket, laptop power supply, printouts
of faces, moving objects, and objects with skin-like texture, such as wood. Figure
7.4 (a) depicts some examples from the training data. Each color frame corresponds
to an evidence type and green frames correspond to face detection, orange frames to heat
detection, red frames to color detection, and yellow frames to motion detection. Note

7.4. Experiments 139
that the training data contained intentionally many cases in which the vision system
produces ambiguous evidences. From this data, MRF models were trained by K-fold
cross-validation, with K = 3.
In order to evaluate the model selection, we reduced the total set of edge types from
32 to 9 since in our case some feature combinations represent the same concept, as
for example, the combinations of heat from the thermo images together with face from
the color images, and face and heat both from the thermo image. The genetic model
selection was evaluated by multiple runs with varied parameter α and varied scoring
metric (Equation 7.4), e.g. based on the false-positive rate, total error, and total area.
In average, the genetic selection yielded the optimal solution after considering 60 ± 7
models, which is more than eight times faster than performing exhaustive search over
all 512 possible models.
For the selection of the final MRF model, we utilized the area-based scoring metric
with α = 2.0. The genetic algorithm selected a classifier which activates, for example,
the edge types motion ∧ f ace, heat ∧ skin, heat ∧ f ace, heat ∧ motion, and forbids
motion ∧ skin, as well as all edge types between the same kind of nodes. Finally, the
selected classifier reached an accuracy of 87.9% at 2.3 ms, in contrast to the classifier
with all edges activated (88.76 % at 12.59 ms) and the classifier with no edges activated
(71.33 % at 1.14 ms). Figure 7.4 shows some examples of the successful application
of the classifier even to hard cases, such as small finger movements, and test persons
completely surrounded by faked evidences.
SVM
MRF
False False Err. False False Err.
Pos. Neg. [%] Pos. Neg. [%]
Human 23 433 39.4 26 143 14.6
Faked 758 0 11.3 151 0 2.3
Both 703 2689 32.8 484 836 12.8
Total 1484 3122 21.0 661 979 7.5
Table 7.2: Comparison of the SVM and MRF classifier: Numbers denote the amount of
wrongly classified evidences in images containing humans, faked evidence, and both.
We compared the performance of the optimized MRF model with a Support Vector
Machine (SVM) [Vapnik, 1995] based classifier. The SVM was trained on the same
features as they were generated for the MRF model, shown in Section 7.2, however,
without encoding of the links between the features. In Table 7.2 the performance for
classifying single evidence frames of both classifiers is reported. The results have been
partitioned into three sets showing the performance on examples containing human evidence,
faked evidence, and both. The result indicates that the optimized MRF model
performs better in terms of false-positive classifications, particularly in situations con-

140 Chapter 7. Victim Detection
taining exclusively faked data.
In the context of Search and Rescue it is desirable to reach a high true-positive rate
on each image, i.e. humans are detected reliably, and a low false-positive rate, i.e. no
victim alarm from faked evidence. This is not directly expressed by the percentage of
correctly classified frames since one wrongly detected frame within an image suffices to
trigger the false alarm, even all the others are detected correctly. Therefore, we counted
the true-positive and false-positive rate for both classifiers image-wise, i.e. images are
counted as true-positive and false-positive if there is a single correct and a single wrong
evidence found, respectively. It turned out that for images containing human evidence
both MRF and SVM reported a victim correctly in 100% of the cases, whereas in images
containing faked evidence, the SVM wrongly reported a victim in 60% and the MRF
in 13% of the cases. Note that the result of the MRF model is comparably good, since
the training data also contained images with more than three different types of faked
evidence at the same time, which makes in this case a distinction from human beings
impossible.
7.5 Related work
Human body detection and tracking has been mainly carried out based on background
subtraction [Wren et al., 1997,Beleznai et al., 2004,Han and Bhanu, 2003] and featurebased
detection [Viola et al., 2003,Cutler and Davis, 2000]. Pfinder is a system for realtime
detection and tracking of human bodies based on background subtraction [Wren
et al., 1997]. The background model, which is continuously updated, utilizes Gaussian
distributions in the YUV color space at each pixel. Humans are modeled by blobs that
are characterized by color, size, and the corresponding Gaussian distributions. It adopts
a Maximum A Posterior (MAP) approach in order to assign pixels either to blobs or
to the background. Beleznai et al. introduced a system for human body detection that
considers the intensity difference between an input frame and a reference image based
on the mean shift procedure [Beleznai et al., 2004]. Their method can be applied in
real-time since all computation is performed on integral images. Han et al. examined
the problem of image registration of both thermal and color images [Han and Bhanu,
2003]. They propose a genetic algorithm-based hierarchical search approach for image
registration, which works on simple body silhouettes generated from background
subtraction of both image types. Viola et al. are using a classifier trained on human
shape and motion features [Viola et al., 2003]. The detector uses images as inputs
and efficiently extracts simple rectangular features using integral images. A cascade of
classifiers is created to achieve superior detection and low false positive rates. Each
stage of the cascade is trained on true and false positives from the previous stage using
Adaboost. Cutler et al. presented a technique for detecting periodic motions patterns,
such as walking [Cutler and Davis, 2000]. They apply time-frequency analysis based

7.6. Conclusion 141
on the short-time Fourier transform (STFT), and autocorrelation for robust periodicity
detection and analysis.
Support Vector Machines (SVMs) have been utilized for detecting human motion
[Cao et al., 2004, Sidenbladh, 2004]. For example, Cao et al. presented a real-time system
for motion detection using SVMs [Cao et al., 2004]. They utilized sets of filtered
images, each encoding a short period of motion history, as input to the SVM. However,
the described system seems not to be suitable for harsh environments, since it requires
human motion to take place in the center of the image and trajectories have to be performed
perpendicular to the optical axis of the camera. MRFs have been successfully
applied to pedestrian tracking [Wu and Yu, 2006] and face detection [Dass et al., 2002].
For example, Dass et al. consider the spatial distribution of gray levels within images
of human faces [Dass et al., 2002]. Feature selection has been examined in various
contexts, e.g. Kohavi et al. provides a good overview on various methods [Kohavi and
John, 1997].
In the context of rescue robotics, Bahadori et al. studied various techniques from
computer vision that have been applied to human body detection [Bahadori and Iocchi,
2003]. Nourbakhsh et al. utilized a sensor fusion approach for incorporating the measurements
from a microphone, IR camera, and conventional CCD camera [Nourbakhsh
et al., 2005]. They assigned to each sensor a confidence value indicating the certainty
of measurements from this sensor and calculated the probability of human presence by
summing over all single sensor observation probabilities, weighted by their confidence
value. However, in contrast to the proposed MRF approach, their method does not
reflect local dependencies of detected evidence in the model.
7.6 Conclusion
We contributed a system that creates MRF models in real-time from motion, color, and
shape evidence, detected by a CCD camera and IR camera, respectively. In order to
reduce computational complexity during inference, the building process of models was
optimized by a genetic algorithm, which decides offline relevant edge types with respect
to the data. Finally, the selected classifier was five times faster than the model
with all edge types activated, while gaining optimal performance in terms of the complexity
trade-off, and near-optimal performance in terms of accuracy. We compared
the optimized model with a SVM and showed that the false-positive rate was significantly
reduced, which is an important aspect when considering victim detection in the
context of rescue robotics. From an image-wise evaluation of the classifier it can be
concluded that the approach reliably detects victims if present and only in hard cases,
i.e. if the number of faked evidences is high, false-positives occur. The classifier performance
could be further improved by introducing temporal relations, i.e. by adding
edges between evidences found in preceding images from the video stream.

142 Chapter 7. Victim Detection
The proposed approach can be considered as a general framework for combining
human evidence detected by sensors. For example, it can easily be extended for incorporating
other types of human evidence, such as audio noise, e.g. tapping, and CO 2
emission. Also given these evidence types, it will be interesting to determine correlations
between them that significantly contribute to the classifier’s performance. Furthermore,
we will consider in future work the extension of the class variable by classes
describing the victim’s state, e.g. aware and unconscious. These classes can also be determined
from correlations between different evidence types. For example, no detection
of motion, but detection of CO 2 emission indicates the victim’s unconscious.

7.6. Conclusion 143
(a) (b) (c)
Figure 7.4: Examples from the test data: (a) Detected evidences: skin color (red), heat
(orange), motion (white), face (green), (b) the same evidences after classification and
(c) all positive evidences clustered into areas.

8 Human In The Loop
8.1 Introduction
One challenge in disaster response is to join local observations made by first responders
in the field, such as blocked or unblocked connections, and found victims, at a central
command post in order to coordinate and schedule search and rescue missions. "Human
in the loop" stands for Multi-Agent-Systems (MAS) that directly combine perceptions
and actions from humans with the decision making of the agents. In fact, humans
are typically connected to the MAS via so-called proxies. Researchers in the field of
Multi-Agent Systems (MAS) developed a rich set of solutions for efficiently coordinating
agents, as well as simulators for evaluating these methods. The RoboCup Rescue
simulation system aims at simulating large-scale disasters and exploring new ways of
autonomous coordination of rescue teams [Kitano et al., 1999]. The goal here is to
provide a software system that reacts to simulated disaster situations by coordinating a
group of simulated agents such as police, ambulance and firefighters. This goal leads
to challenges like the coordination of heterogeneous teams with more than 30 agents,
the exploration of large-scale environments in order to localize victims, as well as the
scheduling of time-critical rescue missions.
Applying the simulation system to the real world is problematic since it lacks the
interface to the real world information. Currently, the simulation system relies on carefully
designed special-purpose map data and observations generated by the simulation
itself, e.g. agent motion is computed by a traffic simulator. However, by using wearable
computing technology, we can provide real-world observations to the simulation system.
This has three uses: First, it can be used to record real-world data from real-world
intervention scenarios and by this, assess the validity of the simulation. Second, it can
be used to observe the reaction of the multi-agent systems to real-world data. Third,
it is a step towards using both the simulation system and the multi-agent systems to
support incident commanders and responders in training and real interventions by providing
autonomous decision support and faster-than-real-time simulation for strategy
decisions.
In this chapter, we propose a wearable computing device for acquiring disaster relevant
data, such as locations of victims and blockades, and show the process of successive
data integration into the RoboCup Rescue simulation system. We assume that the
locations of first responders are automatically tracked, which can either be carried out
145

146 Chapter 8. Human In The Loop
by GPS positioning or PDR combined with RFID-SLAM (see Chapters 3 and 4). Communication
between wearable computing devices and the server is carried out based
on the open GPX protocol [TopoGrafix, 2004] for GPS data exchange, which has been
extended for additional information relevant to the rescue task. By providing examples,
we show how the data can be integrated and how rescue missions can be optimized
by solutions developed on the RoboCup Rescue simulation platform. For this purpose,
we introduce solutions that were developed by the ResQ Freiburg 2004 team [Kleiner
et al., 2004], the winner of RoboCup 2004 ∗ . Besides planning and learning techniques,
this package includes methods for the efficient coordination of victim search, and an
any-time rescue sequence optimization based on a genetic algorithm.
The utilized multi-agent system was extensively evaluated during the Rescue simulation
competition in 2004. The proposed results provide a comparison of the agent
team with other teams, particularly in terms of exploration and mission scheduling.
The efficiency of these methods combined with the demonstrated real-world data integration
indicates that wearable computing technology combined with MAS technology
can serve as a powerful tool for Urban Search and Rescue (USAR). The proposed interface
was demonstrated during the Infrastructure competition at RoboCup 2006, where
it was awarded the first prize [Kleiner et al., 2006a].
The remainder of this chapter is structured as follows. The interface between human
rescue teams and the rescue simulator is proposed in Section 8.2. In Section 8.3 we
introduce agent based solutions for search and rescue that can be applied to collected
real-world data. In Section 8.4 we propose results from experiments on integrating
real-world data into the Rescue simulation system, as well as results from evaluating
the deployed multi-agent system. Finally, we discuss in Section 8.5 related work and
conclude in Section 8.6.
8.2 Interfacing human first responders
In wearable computing, one main goal is to build devices that support users during their
primary task with little or no obstruction. Apart from the usual challenges of wearable
computing [Starner, 2001a, Starner, 2001b], in the case of emergency response, the situation
of the responder is a stressful one. In order to achieve primary task support and
user acceptance, special attention has to be given to user interface design. For this application,
users need the possibility to enter information about their observations, and
need feedback from the system, which recorded and transmitted the information. Furthermore,
the user needs to receive task-related instructions from the command center,
such as assignments to victim search tasks.
The implementation has to cope with multiple unreliable communication systems,
∗ Note that the Open Source version of the software is freely available [ResQ Freiburg, 2004].

8.2. Interfacing human first responders 147
such as existing cell phone networks, special-purpose ad-hoc communication, and existing
emergency-response communication systems. As the analysis of the different
properties of these communication systems is beyond the scope of this chapter, we
will therefore abstract from them and assume an unreliable IP-based connectivity between
the mobile device and a central command post. This assumption is motivated
by the fact that both infrastructure-based mobile communication networks and current
ad-hoc communication systems can transport IP-based user traffic. For example, the
German company IABG implemented HiMoNN (High Mobility Network Node), a mobile
ad-hoc network for catastrophe management allowing rapid information transfer
and secure communications via IP [IABG mbH, 2007].
The situation of the device and its user is also characterized by harsh environmental
conditions related to emergency response, such as fire, smoke, floods, wind, and chemical
hazards. The device has to remain operable under such conditions and additionally
has to provide alternative means of input, e.g. via gestures or textile devices, and output,
e.g. head-mounted display and audio, under conditions that affect human sensing and
action abilities. Moreover, the system has to be integrated into the processes of emergency
response, i.e. having no impact on response times of units and therefore should
be integrated into the existing gear of responders.
The wearable system has to track the location of the person continuously, which
can either be carried out by GPS positioning, if the person walks outdoors, or by the
PDR method described in Chapter 3, if the person walks indoors or close to buildings.
Furthermore, RFID tags, which are already present or which are actively distributed,
have to be detected, and crucial information, such as data on victims, has to be logged
by the system. If communication is possible, recorded data has to be transmitted to a
central command post, which utilizes all trajectories and RFID observations for jointly
optimizing the map by the RFID-SLAM method described in Chapter 4. Figure 8.1
depicts the functional modules of the wearable system.
8.2.1 Preliminary test system
In order to analyze the properties of the communication and localization system, and
to test the software interface to the central server, a preliminary test system has been
implemented, for which three requirements have been dropped: The design criteria according
to harsh environmental conditions, alternative input methods and HMD-based
user interface, and the PDR-based indoor localization capability. Note that the latter
has been already extensively evaluated in Chapter 4.
The communication and localization system is independent of the user requirements
except for the fact that the system has to be portable. Therefore, we chose a mobile
GPS receiver device and a GSM cell phone device as our test implementation platform.
The GPS receiver uses the bluetooth [IEEE, 2002] personal area network standard to
connect to the cell phone. The cell phone firmware includes a Java VM, based on the

148 Chapter 8. Human In The Loop
Figure 8.1: System diagram of the full system based on a GSM phone for data transmission
connected to a belt-worn wearable computer, RFID reader, positioning device
(either GPS or PDR), and a HMD.
J2ME standard with JSR82 extensions, allowing Java applications, which are running
on the VM, to present their user interface on the phone, and to communicate directly
with nearby blue-tooth devices, as well as with Internet hosts via the GSM network
GPRS standard.
The implementation of the test application is straightforward: It regularly decodes
the current geographic position from the NMEA data stream provided by the GPS receiver
and sends this information to the (a priori configured) server IP address of the
central command center. The utilized protocol between the cell phone and the command
center is based on the widely used GPX [TopoGrafix, 2004] standard for GPS
locations. Among other things, the protocol defines data structures for tracks and waypoints.
A track is a time stamped sequence of visited locations, whereas a waypoint
describes a single location of interest, e.g. the peak of a mountain. We extended the
protocol in order to augment waypoint descriptions with information specific to disaster
situations. These extensions allow rescue teams to report the waypoint-relative
locations of RFIDs, road blockades, building fires, and victims. Currently, the wearable
device automatically sends the user’s trajectory to the command center, whereas additional
information, such as the locations of victims or hazards, can be entered manually.
A detailed description of the protocol extension can be found in Appendix 10.1.
8.2.2 Wearable emergency-response system
In order to fulfill all stated requirements, the full system was designed with additional
hard- and software components. The system uses a wearable CPU core, the so-called

8.3. Multi Agent Systems (MAS) for disaster management 149
qbic belt-worn computer [Amft et al., 2004] (see Figure 8.2 (a)). It is based on a ARM
CPU running the Linux operating system, has a bluetooth interface, and can be extended
via USB and RS232 interfaces. The wearable CPU core runs the main application program.
For localization, the same mobile GPS receiver as in the test system is used, but
can be replaced by a non-bluetooth serial device for increased reliability. For communication,
the system can use multiple technologies, the GSM cell phone used for the test
system, can be one of those † .
As already stated, the design of the user interface is crucial for this application.
Therefore, we use a glove as wearable user input device [Lawo et al., 2006] and a
wireless link between the user interface device and the wearable computer. Such an
interface has already been designed for other applications, such as aircraft cabin operation
[Nicolai et al., 2005]. Also the detection of RFIDs is carried out via the glove (see
Figure 8.2 (e)). This facilitates the automatic detection of RFIDs when the user touches
objects, such as a door frame or door knob, which were labeled with RFID tags.
The primary output device is a head-mounted display that can be integrated into
existing emergency-response gear, such as firefighter helmets and masks (see Figure 8.2
(d)). In applications where headgear is not commonly used, the output can be provided
also through a body-worn display device or audio output. The application software
driving the user interface is based on the so-called WUI toolkit [Witt et al., 2006], which
uses an abstract description to define user interface semantics independent of the input
and output devices used. The application code is therefore independent of the devices,
and robust towards different instances of the implementation, i.e. with or without headmounted
display. The WUI toolkit can also take context information into account, such
as the user’s current situation, in order to decide by which device and in what form
output and input are provided.
8.3 Multi Agent Systems (MAS) for disaster
management
As has been shown in Chapter 4, the information collected by the central station can be
used for computing a globally consistent map. This map representation can be utilized
further for the automatic optimization of search and rescue strategies by Multi-Agent
Systems (MAS) methods. The communication protocol described in Section 8.2 has
been particularly tailored according to the protocol of the RoboCup Rescue simulation
kernel. This is motivated by the long-term goal of providing real-world data to the
MAS community, as well as to facilitate the development of robust MAS methods for
† As we assumed IP-based connectivity, flexible infrastructure-independent transport mechanisms such
as MobileIP [Perkins, 2002] can be used to improve reliability over multiple independent and redundant
communication links.

150 Chapter 8. Human In The Loop
(a) (b) (c)
(d)
(e)
Figure 8.2: All parts of the wearable emergency-response system: (a) The qbic beltworn
computer, (b) the Head-Mounted Display (HMD), (c) a firefighter jacket equipped
with an IMU sensor for PDR, (d) the HMD worn by a test person, (e) a wireless RFID
reader device, integrated into a glove.
disaster management. We utilized the code base of the ResQ Freiburg team [Kleiner
et al., 2005a] for evaluating MAS techniques in the context of urban search and rescue.
In the following the integration of data captured by the interface, which has been discussed
in Section 8.2, into the RoboCup Rescue simulation kernel is described. Moreover,
the optimization of search and rescue tasks based on this data is discussed along
with methods for the optimization of rescue sequences.
8.3.1 Real-world data integration
Generally, it is assumed that the wearable device of each human responder continuously
reports the latest recorded positions in form of track messages to the command center,
if communication is possible. Additionally, the rescue team might provide information
for specific locations, as for example, indicating the successful exploration of an area,
or the detection of a victim by triggering the according waypoint message at the current

8.3. Multi Agent Systems (MAS) for disaster management 151
location.
Based on an initial road map, the MAS continuously integrates all collected track
data, and information on the status of roads, e.g. whether they are passable at high costs
or even impassable, as reported by the responders in the field. More specific, each
connection between two locations is augmented with a weight that is computed from
the distance, the amount of blockage, the number of crossings, and the number of other
agents known to be located on the same road. The emerging graph structure is used
as an input for the Dijkstra algorithm [Dijkstra, 1959] computing a connectivity matrix
that stores for each human responder location the distance costs to any other location on
the map. If places are unreachable, costs are infinite. In the worst case, the computation
is in O (m + nlog (n)) for a single responder, where n is the total number of locations
and m is the number of connections between them. The knowledge on the reachability
between locations allows the system to recommend “safe” routes to rescue workers, as
well as to optimize their target assignments by taking travel costs into account. Note that
the connectivity matrix has to be re-computed if new information on the connectivity
between two locations is available, e.g. either if the blockage of a connection has been
observed, or if a formerly blocked connection has been cleared.
In case the procedure is applied with simulated agents, movement interactions are
simulated by the Traffic simulator, and observations are according to a simulated disaster
dynamic carried out by the Collapse simulator and Fire simulator [Nüssle et al.,
2004], which are an integral part of the rescue simulation system. The sequence in Figure
8.3 (a) shows the continuous update of the connectivity matrix for the blue region
in the simulated City of Foligno.
8.3.2 Coordinated victim search
Victim search can only be coordinated efficiently if rescue teams synchronize information
on explored areas. Therefore, the protocol described in Section 8.2 implements
messages for reporting the clearance of explored areas, as well as the detection of a victim.
The command center utilizes this information for efficiently assigning rescue teams
to unexplored locations that are reachable. The sequence in Figure 8.3 (b) shows the
subsequently increasing knowledge on the status of the exploration, being transmitted
to the control center by agents in the field. Regions that are marked with a yellow border
indicate exploration targets assigned by the command center according to a global
exploration strategy, which is described in the following.

152 Chapter 8. Human In The Loop
(a)
(b)
Figure 8.3: Online data integration of information reported by agents exploring the
simulated City of Foligno: (a) The connectivity between the blue region and other locations
increases over time due to the clearance of blocked roads, showing unreachable
locations (white color), and reachable locations (red color). The brighter the red color,
the better the location is reachable. (b) The subsequently increasing knowledge on the
status of the exploration, given the information communicated by agents in the field,
showing passable roads (green color), explored regions with victim detection (green
color) and without (white color). Regions marked with a yellow border are exploration
targets recommended by the command center agent.

8.3. Multi Agent Systems (MAS) for disaster management 153
District exploration.
District exploration is a multi-agent behavior for coordinated victim search. The behavior
guarantees that at any time each agent has an assignment to an unexplored district
on the map. One problem arising when defining districts after a fixed pattern, e.g. by
overlaying equally sized rectangles, is that districts might be differently sized due to
an inhomogeneous distribution of roads and buildings. Furthermore, places within the
same district might not be interconnected, making it unfeasible for agents to entirely
explore the district they were assigned to. Districts are more effective if they are build
according to the building distribution, and according to the connectivity of locations on
the map, e.g. they should consist of locations with a high degree of inter-connectivity.
The connectivity value of two locations depends on the number of alternative paths
that can be found between them, the number of lanes, the degree of blockage of each
connection, and the degree of uncertainty on the state of the connection.
We utilized agglomerative [Bock, 1974] and KD-tree [Bentley, 1975] clustering in
order to calculate a near optimal separation of the map into districts, and to compute
a value expressing the approximative meta-connectivity between them. These methods
calculate from a given connectivity graph G = 〈V, E〉, where V denotes the set of
locations, and E the set of connections between them, a hierarchical clustering. According
to a distance metric, each edge e ∈ E has a weight attached, which is considered
during the clustering procedure. KD-tree clustering is a Divide and Conquer method,
which can be efficiently performed in O ( n log n ) , whereas agglomerative clustering, a
Dynamic Programming method, can be carried out in O ( n 2)‡ [Eppstein, 2000]. However,
in KD-tree clustering the distance metric has to be a L n -norm, e.g. the Euclidean
or Manhattan distance, whereas in agglomerative clustering, the distance metric can be
any arbitrary function.
Agglomerative clustering methods are distinguished by the way how the distance
between two clusters is computed. For example, in Single Linkage the minimum distance
between objects o i and o j from cluster C i and C j is taken as distance between the
clusters:
d ( C i , C j
)
= min
s,t
{
d (s, t) | os ∈ C i , o t ∈ C j
}
, (8.1)
in Complete Linkage the maximum is taken by:
d ( C i , C j
)
= max
s,t
{
d (s, t) | os ∈ C i , o t ∈ C j
}
, (8.2)
and in average linkage the average of all connections between objects from cluster C i
‡ Note that that this bound only applies for Single Linkage, which will be described in the following.

154 Chapter 8. Human In The Loop
and C j by:
d ( C i , C j
)
=
∑
s∈C i
∑t∈C j
d (s, t)
n i n j
(8.3)
We utilized Single Linkage in order to divide city maps into districts for exploration
assignments. One significant advantage of agglomerative clustering is that the state of
a connection, e.g. partial or full blockage, length and width of the road etc., can be
considered directly for building the hierarchy. Figure 8.4 compares KD-tree clustering
with agglomerative clustering on the map of Bremen. The figure shows that blocked
roads (black crosses) are considered during agglomerative clustering, whereas KD-tree
clustering, based on the Euclidean distance metric, separates the city map in nearly
equally sized districts containing places that are unreachable.
(a)
(b)
Figure 8.4: Comparison between (a) KD-tree clustering and (b) agglomerative clustering
on the map of Bremen. Blocked roads (black crosses) are considered during
agglomerative clustering, while KD-tree clustering, based on the Euclidean distance
metric, separates the city map into nearly equally sized districts containing places that
are unreachable from each other.
Since agglomerative clustering represents the true connectivity of places, it is a better
choice for district exploration. The hierarchical clustering, represented by a binary tree,
provides at each level a partitioning of the map into n districts, reflecting the reachability
of locations, e.g. locations which are highly connected are likely to be found within the
same cluster. Given the cluster tree, one can select a tree level according to the size of
the search team, i.e. such that n ≥ m, where m denotes the number of agents and n the
size of the clustering. Due to the fact that blockades on the map and hence the map’s
connectivity is unknown to the agents in advance, the clustering has to be revised each
time new information arrives.

8.3. Multi Agent Systems (MAS) for disaster management 155
Active Exploration.
Exploration within districts can be accelerated significantly if exploration targets are selected
with respects to the expected utility of observed places, e.g. computed from the
agent’s sensor model and traveling costs. Active exploration is a method that considers
exploration utility U and traveling cost C when selecting target locations [Burgard
et al., 2005]. In a search and rescue scenario, the expected utility for visiting places can
be computed by joining observations from agents, such as audio noise and CO 2 concentration,
into a single occupancy grid (see Section 5.2), while considering the senor
model, e.g. maximal detection range and field of view, of each sensor.
We utilized a simplified occupancy grid map for incorporating visual and auditory
sensor observations from multiple agents. Note that also negative information, e.g. no
evidence detection at a certain location, as well as positive information can be incorporated
for increasing the efficiency of the search. Finally, from the set of locations L D
that are within the agent’s district, a target location l t is decided based on the trade-off
between utility U (l) and travel cost T (l):
l t = argmax U (l) − α ∗ T (l) (8.4)
l∈L D
where α is a constant regulating the trade-off between the estimated travel costs and
the exploration utility, which has to be determined experimentally. The estimated travel
costs T (l) are taken from the connectivity matrix, computed by the Dijkstra algorithm.
8.3.3 Rescue sequence optimization
Since time is a critical issue during real disaster response, the survival rate of human
victims depends on the optimal schedule of a rescue mission. For example, after a car
accident on a highway, it is common practice for ambulance teams to decide the sequence
of rescuing victims according to the degree of their injuries and accessibility.
During a large-scale disaster, such as an earthquake, the efficient scheduling of rescue
teams is even more important since there are probably many victims and often an insufficient
number of rescue teams. Furthermore, the time needed for rescuing victims
might significantly vary depending on the collapsed building structures trapping them.
Therefore, it is important that rescue teams inform the station about the state of victims,
e.g. conscious or unconscious, the degree of injuries, and their particular situation, e.g.
surface, trapped, void, or entombed. This information facilitates optimizing the schedule
of rescue missions by the station for maximizing the overall number of survivors.
In RoboCup Rescue, victims are reported with the three variables damage, health and
buridness, expressing an individual’s damage due to fire or debris, the current health that
continuously decreases depending on damage, and the difficulty of rescuing the victim.
From this variables, one can estimate an upper bound for the time necessary to rescue

156 Chapter 8. Human In The Loop
a victim, and a lower bound for the time the victim will survive in the current state.
We utilized Adaptive Boosting (Ada Boost) [Freund and Schapire, 1996] for learning
the regression of a victim’s survival time, and linear regression for predicting the time
needed for performing the rescue task, given the mentioned parameters. Furthermore,
the travel times needed for reaching victim locations have been computed by averaging
over sampled travel times needed by individual agents.
It can be assumed that during a real disaster response scenario expert knowledge can
be acquired for providing rough estimates on the resources needed for victim rescue,
i.e. first responders might be able to estimate whether the removal of debris needs minutes
or hours. Furthermore, the continuous sampling of first responder trajectories, e.g.
acquired by PDR methods or GPS, can be utilized for automatically estimating travel
times for different means of transport, e.g. by car or by feet, between specific locations
[Liao et al., 2007].
Given estimates on the survival time of victims, and the time needed for rescuing
them, it is possible to compute the total number of survivors for any arbitrary rescue
sequence. More specific, one can compute for each rescue sequence S = 〈t 1 , t 2 , ..., t n 〉
of n rescue targets a utility U(S ) that is equal to the number of survivors. Consequently,
the process of determining the optimal rescue sequence that maximizes U(S ) can be
considered as an optimization problem. However, exhaustive search over all possible
n! sequences is already for a little number of targets intractable. The Hungarian
Method [Kuhn, 1955] is a polynomial algorithm for solving scheduling tasks, and finds
the optimal task assignment in O ( mn 2) , where n is the number of workers, and m the
number of tasks. However, the method requires that the time needed until a task is
finished does not influence the overall outcome. This is clearly not the case for rescue
tasks, since victims will not survive if they are rescued too late. Alternatively, Greedy
heuristics can be deployed, for example, to schedule rescue missions first that require
the least amount of time to be accomplished, or to schedule urgent rescue missions
first, e.g. those with the least amount of victim survival time. However, it seems to be
obvious that better solutions might be found by trading-off both of them.
We utilized a Genetic Algorithm (GA) [Holland, 1975] for the online optimization
of rescue sequences. Rescue sequences are represented by integer strings, e.g. S =
〈2, 5, 1, 3, 4〉, encoding each individual solution in a genetic pool. The genetic pool is
initialized with heuristic solutions, for example, sequences that greedily prefer targets
that can be rescued within a short time or urgent targets that have only little chance of
survival. The fitness function of solutions is given by the sequence utility U(S ), which
itself is computed by simulating the expected outcome of the sequence. In order to
guarantee that solutions in the genetic pool are at least as good as the heuristic solutions,
the so-called elitism mechanism, which forces the permanent existence of the best found
solution in the pool, was used. Furthermore, we utilized a simple one-point-crossover
strategy, a uniform mutation probability of p ≈ 1/n, and a population size of 10. One
important property of the proposed method is that it can be executed as an anytime

8.4. Experiments 157
algorithm. At any time the method provides a solution that is at least as good as the
greedy solution, but potentially better with increasing computation time.
8.4 Experiments
8.4.1 Data integration from wearable devices
The integration of real-world data into the RoboCup rescue system was evaluated by
successively processing data transmitted by a test person. The test person was equipped
with the wearable interface described in Section 8.2, while walking several trajectories
in an urban area in the City of Bremen (see Figure 8.5). During this experiment, the
mobile device continuously transmitted the trajectory via GPRS to a central server.
Furthermore, each time the test person had visual contact with a simulated victim § , it
reported via the described GPX protocol a victim found waypoint message to the server.
In order to integrate the data into the rescue system, the received data, encoded by
the extended GPX protocol that represents location by latitude and longitude, was converted
into a grid-based representation. We utilized the Universal Transverse Mercator
(UTM) [Snyder, 1987] projection system, which provides a zone for any location on
the surface of the Earth, where coordinates are described relatively to this zone (see
Section 2.2.3). By calibrating maps from the rescue system to the point of origin of the
UTM coordinate system, locations from the GPS device can directly be mapped. In order
to cope with erroneous data, we decided to simply ignore outliers, i.e. locations far
from the track, that were detected based on assumptions made on the test person’s maximal
velocity. Figure 8.5 (b) shows the successive integration of the received data into
the rescue system and Figure 8.5 (a) displays the same data plotted by GoogleEarth.
Note that GPX data can be displayed directly by GoogleEarth without requiring any
pre-processing.
8.4.2 Coordinated victim search
The efficiency of the proposed method for exploration and victim search was evaluated
by analyzing log files recorded during the RoboCup Rescue simulation competition
in 2004. During this competition, 20 international teams competed on different maps
simulated by the rescue system. The methods described in this chapter were an integral
part of the ResQ Freiburg team which achieved the first place of the competition. Other
teams applied, for example, Reinforcement Learning (DAMAS-Rescue and RoboAkut),
and MAS group forming methods (NITRescue04).
During emergency response, the newest information is the most valuable. For example,
the earlier victim locations are known, the better their rescue can be scheduled.
§ Note that arbitrary locations in Bremen were selected as victim locations and were visually marked.

158 Chapter 8. Human In The Loop
(a)
(b)
Figure 8.5: Successive integration of data reported by a test person equipped with
a wearable device. (a) The real trajectory and observations of victims plotted with
GoogleEarth (victims are labeled with “civFound”). (b) The same data integrated into
the rescue system (green roads are known to be passable, white buildings are known as
explored, and green dots indicate observed victims).

8.4. Experiments 159
Figure 8.6 shows the number of civilians found during each cycle on the RandomMap
during the final and semi-final, respectively. The results confirm the efficiency of the
90
100
80
90
70
80
Percent Civilians Found
60
50
40
30
20
10
0
0
150
Time
(a)
ARK
BAM
Caspian
DAMAS
ResQFreiburg
SBCe
SOS
TheBlackSheep
300
Percent Civilians Found
70
60
50
40
30
20
10
0
0
150
Time
(b)
BAM
Caspian
DAMAS
ResQFreiburg
300
Figure 8.6: The number of civilians found by exploration on a randomly generated map
during the (a) final and (b) semi-final.
exploration strategy of ResQ Freiburg. At any given time our agents knew about more
civilians than the agents of any other team.
Table 8.1 shows the percentage of buildings that were searched by the agents , and
Table 8.2 ‖ , shows the percentage of found civilians during each run. Results are shown
from all rounds of all teams that passed the preliminaries during the competition. All
values are according to the state of the world at the last round, e.g. the percentage of
explored buildings at round 300. Bold numbers denote the best results that have been
achieved during the respective rounds. The results show that Caspian explored most of
the buildings. However, ResQ Freiburg was able to find more victims, although they
explored less buildings, which indicates a higher degree of exploration efficiency.
8.4.3 Victim rescue sequence optimization
In order to evaluate the performance of the online genetic algorithm, we executed multiple
simulation runs, once with Greedy optimization and once with GA-based optimization
for scheduling rescue missions. Figure 8.7 depicts the difference between a greedy
rescue target selection, i.e. to prefer targets that can be rescued fast, and the GA-based
selection optimization. It can be seen that GA-based optimization clearly increases the
number of rescued civilians, even the time for making decisions is limited. Within each
Note that full communication of visited locations as well as exploitation of a sensor model was assumed.
‖ Note that civilians are considered as being found if one of the agents was within their visual range.

160 Chapter 8. Human In The Loop
ResQ Damas Caspian BAM SOS SBC ARK B.Sheep
Final-VC 83,48 83,24 87,02 67,27 N/A N/A N/A N/A
Final-Random 69,62 72,62 78,13 49,92 N/A N/A N/A N/A
Final-Kobe 89,19 92,97 89,73 94,19 N/A N/A N/A N/A
Final-Foligno 84,15 85,25 86,73 74,29 N/A N/A N/A N/A
Semi-VC 69,39 72,86 77,42 45,08 52,01 52,87 47,92 59,72
Semi-Random 78,91 68,73 71,91 54,36 59,36 70,27 46,18 46,18
Semi-Kobe 85,41 96,22 92,97 95,54 66,62 97,30 99,46 91,89
Semi-Foligno 74,75 89,12 84,98 62,49 65,35 92,53 79,08 20,74
Round2-Kobe 87,16 90,68 95,00 91,76 80,54 94,19 99,46 92,43
Round2-Random 81,18 80,94 88,61 84,53 60,67 94,24 82,61 87,89
Round2-VC 83,40 70,18 84,58 40,44 67,74 87,88 N/A 89,54
Round1-Kobe 87,43 90,27 94,05 96,08 96,62 97,70 97,84 80,95
Round1-VC 85,37 90,48 95,28 94,26 N/A 97,72 100,00 91,35
Round1-Foligno 83,78 90,05 90,05 60,00 54,65 88,57 67,37 13,00
Number of Wins: 1 1 4 1 0 2 4 1
AVG %: 81,66 83,83 86,89 72,16 67,06 87,33 79,99 67,37
STD %: 5,82 9,98 7,87 22,21 13,87 14,59 21,84 30,80
Table 8.1: Percentage of explored buildings.
ResQ Damas Caspian BAM SOS SBC ARK B.Sheep
Final-VC 97,22 94,44 100,00 81,94 N/A N/A N/A N/A
Final-Random 90,91 85,71 81,82 70,13 N/A N/A N/A N/A
Final-Kobe 98,77 97,53 95,06 98,77 N/A N/A N/A N/A
Final-Foligno 96,67 96,67 96,67 72,22 N/A N/A N/A N/A
Semi-VC 77,92 77,92 85,71 45,45 53,25 53,25 50,65 63,64
Semi-Random 88,51 73,56 72,41 63,22 67,82 80,46 52,87 55,17
Semi-Kobe 100,00 100,00 100,00 98,61 79,17 100,00 100,00 97,22
Semi-Foligno 90,12 95,06 86,42 81,48 83,95 97,53 85,19 30,86
Round2-Kobe 98,89 98,89 97,78 95,56 91,11 100,00 100,00 98,89
Round2-Random 98,89 95,56 98,89 81,11 70,00 96,67 85,56 94,44
Round2-VC 92,22 78,89 90,00 45,56 72,22 88,89 N/A 87,78
Round1-Kobe 94,29 100,00 100,00 98,57 100,00 100,00 94,29 78,57
Round1-VC 100,00 100,00 100,00 97,14 N/A 100,00 100,00 98,57
Round1-Foligno 100,00 97,14 94,29 77,14 74,29 92,86 77,14 14,29
Number of Wins: 9 4 7 1 1 5 3 0
AVG %: 94,60 92,24 92,79 79,06 76,87 90,97 82,85 71,94
STD %: 7,17 10,53 9,03 20,75 13,73 14,69 19,35 30,25
Table 8.2: Percentage of found civilians.
minute, the GA was able to compute approximately 300, 000 solutions on a 1.0GHz
Pentium4 computer.
Finally, Table 8.3 shows the number of civilians saved by each team: ResQ Freiburg
saved more than 620 civilians during all rounds, which are 35 more than the second best
and 59 more than the third best in the competition.

8.4. Experiments 161
100
Greedy-Heuristic
Genetic Algorithm
90
# Civilians
80
70
60
50
0 1 2 3 4 5 6 7 8 9
Map
Figure 8.7: The number of civilian survivors after applying Greedy rescue sequence
optimization and GA-based sequence optimization in different simulated cities.
ResQ Damas Caspian BAM SOS SBC ARK B.Sheep
Final-VC 42 43 52 34 N/A N/A N/A N/A
Final-Random 32 25 29 16 N/A N/A N/A N/A
Final-Kobe 46 45 46 30 N/A N/A N/A N/A
Final-Foligno 66 54 50 29 N/A N/A N/A N/A
Semi-VC 18 15 17 12 11 12 12 14
Semi-Random 22 26 16 14 20 14 15 15
Semi-Kobe 57 47 54 52 20 39 34 44
Semi-Foligno 37 46 44 43 42 28 29 24
Round2-Kobe 57 37 43 50 43 35 28 43
Round2-Random 52 48 39 45 47 44 50 37
Round2-VC 31 33 32 24 37 51 N/A 34
Round1-Kobe 45 51 47 43 47 31 25 34
Round1-VC 62 62 55 57 N/A 51 54 44
Round1-Foligno 53 53 37 33 37 41 30 23
#Wins: 9 5 2 0 0 1 0 0
Σ TOTAL: 620 585 561 482 304 346 277 312
Σ SEMI+PREM 434 418 384 373 304 346 277 312
Table 8.3: Number of saved civilians.

162 Chapter 8. Human In The Loop
8.5 Related Work
Nourbakhsh and colleagues utilized the MAS Retsina for mixing real-world and simulationbased
testing in the context of Urban Search and Rescue [Nourbakhsh et al., 2005].
Wickler et al. introduced To Do lists within the constraint model to coordinate
activities in an emergency response scenario [Wickler et al., 2006]. They utilize a Hierarchical
Task Network (HTN) planner that can be used to synthesize courses of action
automatically. Their system has also been used by Siebra and Tate to support coordination
of rescue agents in the RoboCup Rescue simulation [Siebra and Tate, 2005].
Oh et al. applied agent technologies in post-disaster planning contexts by an agentbased
decision support system that can identify good candidate locations for temporary
housing [Oh et al., 2006]. Schurr and colleagues [Schurr et al., 2005] introduced the
DEFACTO system, which enables agent-human cooperation and was evaluated in the
firefighting domain with the RoboCup Rescue simulation package. U et al. proposed
a generalized agent model for the RoboCup Rescue simulation system [U and Reed.,
2006]. Their extension facilitates agents that have multiple roles, and can change their
capabilities during a simulation, enabling a richer mix of agent behaviors.
Takahashi discusses the practical usage of disaster management systems for local
governments by an evaluation of RoboCup Rescue simulation data, with the intention
of guiding future research topics towards the development of practical disaster simulation
systems [Takahashi, 2006]. Noda and Meguro proposed the standard protocol
MISP and a simple database system DaRuMa for the purpose of collecting and sharing
disaster information about damaged areas for supporting decision-making in rescue
processes [Noda and Meguro, 2006]. Kuwata et al. introduced a simulation framework,
based on RoboCup Rescue simulation, named RoboCupRescue Human-In-The-
Loop Agent Simulation (RCR-HITLAS), in which humans can act as agents with various
roles. [Kuwata et al., 2006]. The purpose of the system is to measure the performance
of user interfaces and decision support systems used by humans that humans, as well as
their skills.
8.6 Conclusion
We introduced the preliminary design of a wearable computing device in conjunction
with a multi-agent based disaster simulator, which both can be utilized for optimizing
USAR missions. It has been generally demonstrated that the system is capable of integrating
trajectories and observations captured during real-world experiments into the
RoboCup Rescue simulation platform. As shown by the results from simulation runs,
the integrated data can serve as a basis for coordinating exploration, e.g. by directing
teams to unexplored regions, as well for scheduling rescue missions. The utilized MAS
approaches were evaluated during various simulation runs, and have also been com-

8.6. Conclusion 163
pared with strategies developed by other teams during the RoboCup Rescue simulation
competition in 2004.
The proposed interface can be extended further by additional sensors, such as CO 2
detectors, temperature sensors, and cameras, for facilitating the transmission of images
and data that document the degree of damages or the state of victims at certain locations.
The overall goal in this context is to design as system that integrates all observations
from both human responders and robot rescue teams, while reducing the burden, i.e.
the cognitive load, of the user. The strategic overview of the scenario allows incident
commanders to efficiently combine the capabilities of individual team members by developing
multi-robot and multi-human plans, given the available resources.
In Chapter 4 the automatic fusion of maps generated by robot and humans has been
demonstrated. In future research, we intend to extent this approach by integrating features
extracted from local observations taken by wearable devices or sensors mounted
on robots. Furthermore, we will consider the evaluation of the coordination mechanism
in real scenarios, e.g. to guide first responders via the wearable computing device during
an exercise by automatically generated plans. It would be particularly interesting
to compare this approach with conventional methods as they are used in emergency
response nowadays, i.e. to determine the impact of agent technology in time critical
missions.
Future work in the Rescue simulation league aims at opening the system to resources
available on the Internet, such as map data from Google Maps, and any other freely
available location-specific data. This direction has the primary goal to close the loop
between disaster reality and disaster simulation.

9 Summary And Future Work
To transfer robot technology from laboratory experiments into the real world for solving
socially significant problems requires research and development being undertaken close
to reality. An exchange of practical knowledge with experts from the field, such as
emergency response personnel, is indispensable. Only their experience can facilitate the
development of deployable autonomous robot platforms, as well as their integration into
human operation. Performance metrics, such as benchmark arenas, are one promising
approach to this goal.
Current benchmarking scenarios do clearly not yet capture the whole magnitude of
problems that robots encounter after a real disaster. In this scenario, mapping methods
are required that also function when robots operate in confined spaces within collapsed
structures, climb-up walls, or fly autonomously over the terrain. The idea behind benchmarking
autonomous robot systems in USAR (Urban Search And Rescue) scenarios is
to continuously increase the level of difficulty year by year in order to promote stepwise
research in this field. Up to now, methods from autonomous robotics for USAR are just
at their advent and robots that have been deployed after a real disaster were mainly
tele-operated by human beings.
Lessons learned from early stage performance metrics already showed that assumptions
holding under mild conditions, such as office-like environments, do not necessarily
hold under harsh conditions. For example, conventional SLAM methods assume a constant
growth of the positioning error during pose tracking, e.g. the longer the track the
larger the variance. Unstructured environments have the property that conditions can
arbitrarily change, e.g. the underground might be smooth floor, slippery, or even rough
terrain, requiring robots to be aware of their context. Data association requires robust
visual features in order to function reliably, which are not constantly found under those
conditions. Furthermore, robotic methods must not interfere with task-specific goals
of operations in the field. For example, conventional mapping algorithms require the
repeated visiting of places in order to refine the map by loop closure, which turns out to
be a suboptimal behavior during victim search and exploration.
Finally, the deployment of robotics technology is only helpful if it complements human
task execution. Here the ultimate goal to is to optimally coordinate human and
robot rescue teams, i.e. to distribute tasks among them according to their individual
skills. Therefore, methods have to address data exchange issues directly, such as the
exchange of maps and observations, while coping with communication perturbations,
or even temporarily loss of communication.
165

166 Chapter 9. Summary And Future Work
In the light of these considerations three core problems were addressed in this thesis,
which are localization, mapping, and exploration with a strong focus on human
integration and teamwork of both robots and humans.
9.1 Summary
In Chapter 2 hardware for Rescue Robotics was introduced and design criteria for robot
platforms built for the deployment in emergency response scenarios were discussed.
According to these criteria, two robot platforms were contributed that have been built
for the development and evaluation of algorithms presented in this thesis. Furthermore,
different sensor types for localization, mapping, and victim search were discussed.
The problem of tracking humans and robots was subject of Chapter 3. A novel
method for slippage sensitive pose tracking on wheeled robot platforms and a novel
method for visual odometry on tracked platforms were contributed. While these methods
were particularly designed for two specific application scenarios, which are the
rapid mapping of a large-scale environment by wheeled robots, and the mapping of
rough terrain by tracked robots, they can basically be considered as independent modules
that can be used for different tasks. They were intensively tested under harsh environmental
conditions, and evaluated within USAR benchmark scenarios, such as the
RoboCup Rescue autonomy competition. Furthermore, an existing method for tracking
the pose of humans was evaluated and discussed. Results showed that human pose
tracking is feasible over larger distances, accompanied with increasing positioning errors.
In Chapter 4 RFID-SLAM, a novel method for SLAM, was contributed. This method
allows the efficient generation of globally consistent maps, even if the density of landmarks
is comparably low. For example, the method corrected an outdoor large-scale
map within a few seconds from odometry data and RFID perceptions only. This has
been partially achieved due to reliable pose tracking based on slippage sensitive odometry,
but also due to the data association solved by RFIDs. Solving data association
by RFIDs allows to speed-up the route graph corrections by decomposing the problem
into optimization and interpolation. Furthermore, we demonstrated the joint correction
of trajectories from human and robot teams. The joint correction of robot and human
trajectories is of major importance since it is very likely that robots and humans explore
different areas during emergency response. For example, robots will be deployed in
hazardous places or places lacking of air, which humans cannot access. We showed that
RFID-SLAM allows to improve maps by loop closure without requiring humans and
robots to perform loops while executing their primary task. Due to the joining of routes
via RFID connection points, loops automatically emerge. This is a necessary requirement
if applying SLAM in USAR situations. In such situations, emergency responders
have clear goals that have to be accomplished within a short amount of time, for exam-

9.1. Summary 167
ple, they have to explore the environment for victims, and therefore intentionally try to
avoid to visit places repeatedly.
Finally, a decentralized version of RFID-SLAM that utilizes the memory capacity
of RFIDs was contributed. The results show that sharing information between single
agents allows to optimize globally their individual paths, even if they are not able to
communicate directly. This has the advantage, particularly during disaster response,
that this method can be applied if communication is disturbed by building debris and
radiation. If communication is temporarily available, the process can be accelerated by
exchanging data between the agents directly.
In the disaster response scenario RFID-SLAM offers other practical advantages. For
example, humans can communicate with RFIDs via a PDA and leave behind information
related to the search, such as discovered victims or hazardous places. The idea
of labeling locations with information that is important to the rescue task has already
been applied in the past. During the disaster relief in New Orleans in 2005, rescue task
forces marked buildings with information concerning, for example, hazardous materials
or victims inside the buildings [FEMA, 2003]. The RFID-based marking of locations
is a straight forward extension of this concept.
In Chapter 5 a method for the building of elevation maps from readings of a tilted
LRF was contributed. In contrast to former work, the proposed method integrates all
six degrees-of-freedoms of the robot in real-time, allowing the mapping of rough terrain
while the robot navigates over obstacles such as ramps. The quantitative evaluation
of conducted indoor and outdoor experiments, partially in testing arenas proposed by
NIST for USAR, showed that the proposed method can be deployed in real-time while
leading to a robust mapping of the environment. The result was supported by the visual
odometry method for pose tracking, which significantly improved the accuracy of scan
matching. As we showed in another work, the generated elevation maps can be utilized
for structure classification and the planning of skill execution [Dornhege and Kleiner,
2007b].
In Chapter 6 a novel method for coordinated exploration of large robot teams was
contributed. The approach, which is based on RFID technology for indirect communication,
is composed of two parts. First, distributed local search, with the notable properties
of not requiring direct communication and scaling with the number of agents.
Second, global task assignment and multi-robot path planning for monitoring the local
exploration and restarting it at better locations. A novel graph structure based on RFIDs
was introduced, which allows a significantly smaller representation of the environment
compared to grid based approaches. The experimental results from RoboCup showed
that RFID-based coordination scales-up with large robot teams exploring large environments.
Moreover, the global task assignment based on genetic sequence optimization
was evaluated. We showed that this method allows to improve the performance of the
local approach for robot teams consisting of up to 20 robots. Due to a unified exchange
of data, e.g. as has been shown with RFID-SLAM, the method can also be extended for

168 Chapter 9. Summary And Future Work
integrating humans into the search.
The problem of real-time detection of victims in disaster areas was discussed in Chapter
7. In order to scale-up the detection rate of weak classifiers, a system for creating
optimized MRF models was contributed, which computes offline relevant feature combinations
with respect to the data. The optimized classifier was five times faster than
the model with all feature combinations activated, while gaining optimal performance
in terms of the complexity trade-off, and near-optimal performance in terms of accuracy.
We compared the optimized model with a SVM classifier and showed that the
false-positive rate was significantly reduced, which is an important aspect when considering
victim detection in the context of emergency response. The proposed approach
can be considered as a general framework for combining human evidence detected by
sensors. For example, it can easily be extended for incorporating other types of human
evidence, such as audio noise, e.g. tapping and CO 2 emission.
In the last chapter, Chapter 8, the preliminary design of a wearable computing device
in conjunction with a multi-agent based disaster simulator for optimizing USAR
mission scheduling, was contributed. It was generally demonstrated that the system is
capable of integrating trajectories and observations captured during real-world experiments
into the RoboCup Rescue simulation platform. As shown by the results from
simulation runs, the integrated data can serve as a basis for coordinating exploration,
e.g. by directing teams to unexplored regions as well as for scheduling rescue missions.
The proposed interface can be extended further by additional sensors, such as
CO 2 detectors, temperature sensors, and cameras, for facilitating the transmission of
images and data that document the degree of damages or the state of victims at certain
locations. The overall goal in this context is to design a system that integrates all observations
from both human responders and robot rescue teams at an incident commander
perspective. The strategic overview of the scenario allows incident commanders to efficiently
combine the capabilities of individual team members by developing multi-robot
and multi-human plans, given the available resources.
9.2 Future Work
As results from PDR experiments showed, the tracking of human beings based on accelerometers
leads to promising results. However, also these results are only preliminary
compared to the situations first responders are exposed to while rescuing victims
after a real disaster. They might crouch, run, and climb within very different and individual
modes, generating a wide range of ambiguous sensor values. Further research
has to be undertaken in order to capture the whole range of human motion.
The proposed method for RFID-SLAM has shown to work robustly in various scenarios.
However, one drawback of this method is given by the low-range detection of
RFIDs due to the utilized technology. In future work, we will evaluate RFID technology

9.2. Future Work 169
operating in the UHF frequency domain, allowing reading and writing within distances
of meters, and to extend the proposed approach accordingly.
The approach of RFID technology-based coordination has been tested mainly in simulated
disaster areas. In future work, it will be evaluated on a team of real robots
equipped with long-range RFID detectors while exploring an outdoor area. Furthermore,
the plan execution phase of the approach will be improved. It is considered to
represent the generated multi-robot plans by Petri nets, allowing to verify online plans
generated from task assignment and path planning. Given this representation, parallel
plan executions can be simulated beforehand for predicting unsafe situations. On
the one hand, this will facilitate the optimization of robot waiting times if they are involved
within conflicts. On the other hand, it will enable deadlock detection during the
planning phase.
The proposed victim detection method serves as a generalized framework and hence
can be further extended. In future work we will consider to extend the binary class
variable by classes describing the victim’s state, such as aware and unconscious. These
classes can also be determined from correlations between different evidence types. For
example, the detection of CO 2 emission and heat without the detection of motion can
indicate an unconscious victim.
Most importantly, future work will deal with the problem of combining RFID route
graph optimization with local elevation mapping in order to facilitate the building of
large and consistent elevation maps in real-time. Here the basic idea is to anchor locally
generated elevation maps within the corrected graph. These map patches can then
be loaded into memory if they are within range of the robot’s current location. Furthermore,
the saved data can be utilized offline for generating a global elevation map
representation of the environment. This representation opens the door for solving realtime
mapping and navigation in large-scale environments, which is an indispensable
requirement for future projects and benchmarks. For example, the 600.000US $ endowed
TechX challenge, a competition organized by the Singaporean national Defense,
Science and Technology Authority (DSTA), requires a one hour fully autonomous exploration
of a larger area containing buildings and rough outdoor terrain. Robots have
to locate targets in buildings and return to their starting location. Such benchmarks demand
highly sophisticated robot systems fulfilling their mission robustly during a long
time operation, while having to rely on themselves completely.
It can be expected for the future that rescue robots continuously improve according
to the increasing difficulty of benchmarking scenarios. Moreover, it is planned by
the RoboCup Rescue community to introduce standardized map representations and
communication protocols, which will be an important milestone towards implementing
heterogeneous robot teams. A common format could serve not only for comparisons
between different approaches, but also as a useful tool to exchange spatial information
between heterogeneous agents, as well as to compare more rigorously results coming
from simulation and real world systems. Furthermore, it makes it possible to provide

170 Chapter 9. Summary And Future Work
floor plans of building structures to rescue teams in advance, as is common practice
during real rescue missions. Here, the long term goal is to develop communication
standards in the simulation league and to transfer them to the robot league, after they
have been extensively tested.

10 Appendix
10.1 Rescue communication protocol
This type describes an extension of GPX 1.1 waypoints.
Waypoints in the disaster area can be augmented
with additional information, such as observations of fires,
blockades, victims, and RFIDs.
This type describes information on a victim
relatively to the waypoint.

172 Chapter 10. Appendix
This type describes the observation of an RFID tag, which can
further be used by the station to correct the agents trajectory.
This type describes detected road blockages
relatively to the waypoint.

10.1. Rescue communication protocol 173

174 Chapter 10. Appendix
This type contains a distance value measured in meters.
This type contains a bearing value measured in degree.

SHEET 1 OF 1
10.2. Construction drawings 175
10.2 Construction drawings
10.2.1 3D Laser Range Finder (LRF)
Figure 10.1: Perspective drawing of the 3D LRF mechanics.
4
52
M4 durch,
mit Gewinden
3
43
41
22
28
36,60
43,35
53,35
7
49
56
5
51
56
4xM4 ohne Gewinde
5
95
100
NAME
DIMENSIONS ARE IN MM
TOLERANCES:
DRAWN
FRACTIONAL
CHECKED
ANGULAR: MACH BEND
TWO PLACE DECIMAL
ENG APPR.
THREE PLACE DECIMAL
PROPRIETARY AND CONFIDENTIAL
MFG APPR.
THE INFORMATION CONTAINED IN THIS
MATERIAL
Q.A.
DRAWING IS THE SOLE PROPERTY OF
--
COMMENTS:
. ANY
REPRODUCTION IN PART OR AS A WHOLE
FINISH
NEXT ASSY USED ON
--
WITHOUT THE WRITTEN PERMISSION OF
IS
PROHIBITED.
APPLICATION DO NOT SCALE DRAWING
DATE
SIZE
A
Institut für Informatik
T-Stück für Servo
Anzahl:2
SCALE:1:1
DWG. NO.
WEIGHT:
REV.
Figure 10.2: Mechanical drawing 3D LRF: servo mount.

176 Chapter 10. Appendix
Figure 10.3: Mechanical drawing 3D LRF: U-shaped piece.
10.2.2 Zerg robot
Figure 10.4: Perspective drawings Zerg robot mechanics.

10.2. Construction drawings 177
Figure 10.5: Mechanical drawing Zerg robot: inner support frame.
Figure 10.6: Mechanical drawing Zerg robot: front part.

178 Chapter 10. Appendix
Figure 10.7: Mechanical drawing Zerg robot: bottom part.
Figure 10.8: Mechanical drawing Zerg robot: side part.

10.2. Construction drawings 179
Figure 10.9: Mechanical drawing Zerg robot: back part.
Figure 10.10: Mechanical drawing Zerg robot: cover back part.

180 Chapter 10. Appendix
Figure 10.11: Mechanical drawing Zerg robot: cover front part.
Figure 10.12: Mechanical drawing Zerg robot: assembling cube.

10.2. Construction drawings 181
10.2.3 RFID tag releaser
Figure 10.13: Perspective drawing of the RFID tag releaser.

182 Chapter 10. Appendix
Figure 10.14: Mechanical drawing RFID tag releaser: bottom back part.
Figure 10.15: Mechanical drawing RFID tag releaser: bottom left part.

10.2. Construction drawings 183
Figure 10.16: Mechanical drawing RFID tag releaser: bottom right part.
2,50
40
2,50
17
15
2
40
UNLESS OTHERWISE SPECIFIED:
NAME
DATE
PROPRIETARY AND CONFIDENTIAL
THE INFORMATION CONTAINED IN THIS
DRAWING IS THE SOLE PROPERTY OF
. ANY
REPRODUCTION IN PART OR AS A WHOLE
WITHOUT THE WRITTEN PERMISSION OF
IS
PROHIBITED.
NEXT ASSY
APPLICATION
USED ON
DIMENSIONS ARE IN MM
TOLERANCES:
FRACTIONAL
ANGULAR: MACH BEND
TWO PLACE DECIMAL
THREE PLACE DECIMAL
INTERPRET GEOMETRIC
TOLERANCING PER:
MATERIAL
FINISH
DRAWN
CHECKED
ENG APPR.
MFG APPR.
Q.A.
COMMENTS:
TITLE:
Schieber_Oben
SIZE
A
DWG. NO.
DO NOT SCALE DRAWING SCALE: 2:1 WEIGHT: SHEET 1 OF 1
REV
5 4 3 2 1
Figure 10.17: Mechanical drawing RFID tag releaser: slider upper part.

184 Chapter 10. Appendix
Figure 10.18: Mechanical drawing RFID tag releaser: slider lower part.
Dieses Teil wird 2-mal benötigt!
5
13,25
20
2,50
5
13
3
A
A
13
8
5
10
2
Gewinde für M5
durchgehend.
Gewinde für M2
10 mm tief.
SCHNITTDARSTELLUNG A-A
MAßSTAB 4 : 1
UNLESS OTHERWISE SPECIFIED:
NAME
DATE
PROPRIETARY AND CONFIDENTIAL
THE INFORMATION CONTAINED IN THIS
DRAWING IS THE SOLE PROPERTY OF
. ANY
REPRODUCTION IN PART OR AS A WHOLE
WITHOUT THE WRITTEN PERMISSION OF
IS
PROHIBITED.
NEXT ASSY
APPLICATION
USED ON
DIMENSIONS ARE IN MM
TOLERANCES:
FRACTIONAL
ANGULAR: MACH BEND
TWO PLACE DECIMAL
THREE PLACE DECIMAL
INTERPRET GEOMETRIC
TOLERANCING PER:
MATERIAL
FINISH
DRAWN
CHECKED
ENG APPR.
MFG APPR.
Q.A.
COMMENTS:
TITLE:
Servohalter (2x)
SIZE
A
DWG. NO.
DO NOT SCALE DRAWING SCALE: 5:1 WEIGHT: SHEET 1 OF 1
REV
5 4 3 2 1
Figure 10.19: Mechanical drawing RFID tag releaser: servo mount.

10.2. Construction drawings 185
.
Figure 10.20: Mechanical drawing RFID tag releaser: RFID tag container.

186 Chapter 10. Appendix
10.3 Schematics
(a)
(b)
Figure 10.21: Schematics of (a) the Power Supply Board, and (b) the Sensor board of
the Zerg robot.

10.3. Schematics 187
Figure 10.22: Schematics of the Micro Controller Board of the Zerg robot.

188 Chapter 10. Appendix
Figure 10.23: Schematics of the Micro Controller Board of the Lurker robot.

10.3. Schematics 189
Figure 10.24: Schematics of the Relais board of the Lurker robot.

List of Figures
1.1 First responders and destroyed building structures . . . . . . . . . . . . 2
1.2 Disaster mitigation statistics . . . . . . . . . . . . . . . . . . . . . . . 3
2.1 Evaluating Rescue Robots I . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2 Evaluating Rescue Robots II . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Legged rescue robots . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Inertial Measurement Unit (IMU) . . . . . . . . . . . . . . . . . . . . 19
2.5 UTM map of Europe . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.6 CCD cameras and IR camera . . . . . . . . . . . . . . . . . . . . . . . 24
2.7 A hand-crafted light-weight 3D Laser Range Finder . . . . . . . . . . . 26
2.8 The Zerg robot measuring audio noise and CO 2 . . . . . . . . . . . . . 27
2.9 Sound source detection by the Crosspower Spectrum Phase . . . . . . . 27
2.10 Robots of the team Rescue Robots Freiburg . . . . . . . . . . . . . . . 32
2.11 Human Robot Interaction (HRI) - RoboGui . . . . . . . . . . . . . . . 35
2.12 Human Robot Interaction (HRI) - IncidentCommander . . . . . . . . . 35
3.1 Dead reckoning on a skid-steering 4WD robot driving in a cellar . . . . 39
3.2 Scan matching with dead reckoning on a skid-steering 4WD robot . . . 41
3.3 Slip detection on a 4WD robot . . . . . . . . . . . . . . . . . . . . . . 42
3.4 KLT feature tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.5 Example of the Monte Carlo algorithm . . . . . . . . . . . . . . . . . . 46
3.6 Acceleration patterns during human walking . . . . . . . . . . . . . . . 50
3.7 Test person with Xsens MTi and Holux GPS device . . . . . . . . . . . 51
3.8 Conventional odometry compared to slippage sensitive odometry . . . . 53
3.9 Zerg robot during the final of the Best in Class autonomy competition . 53
3.10 Accumulating distance error of the visual odometry . . . . . . . . . . . 55
3.11 Distance of the visual odometry method compared to ground truth . . . 56
3.12 Lurker robot performing vision-based pose tracking . . . . . . . . . . . 57
3.13 Comparing elevation mapping methods . . . . . . . . . . . . . . . . . 58
3.14 Comparison between PDR (green trajectory) with GPS ground truth . . 60
4.1 The spring-mass analogy of RFID graphs . . . . . . . . . . . . . . . . 69
4.2 Decentralized SLAM of three agents . . . . . . . . . . . . . . . . . . . 76
4.3 RFID tag release mechanism . . . . . . . . . . . . . . . . . . . . . . . 79
191

192 List of Figures
4.4 RFID-SLAM at different levels of noise . . . . . . . . . . . . . . . . . 80
4.5 RFID-SLAM mapping of a cellar . . . . . . . . . . . . . . . . . . . . . 81
4.6 RFID-SLAM for mapping buildings . . . . . . . . . . . . . . . . . . . 82
4.7 Pose errors from RFID-SLAM with varying number of RFIDs . . . . . 83
4.8 Covariance bounds of RFID-SLAM in large-scale environments . . . . 84
4.9 RFID-SLAM in large-scale environments . . . . . . . . . . . . . . . . 84
4.10 RFID-SLAM with pedestrians semi-indoors . . . . . . . . . . . . . . . 86
4.11 Pose errors from RFID-SLAM with pedestrians semi-indoor . . . . . . 86
4.12 RFID-SLAM in urban environments . . . . . . . . . . . . . . . . . . . 87
4.13 Pose errors from RFID-SLAM in urban environments . . . . . . . . . . 88
4.14 Decentralized SLAM experiment . . . . . . . . . . . . . . . . . . . . . 89
4.15 Human-robot SLAM . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.1 Occupancy grid map of an office-like environment . . . . . . . . . . . . 94
5.2 Transforming range measurements to height values . . . . . . . . . . . 96
5.3 Dead reckoning for building elevation maps . . . . . . . . . . . . . . . 100
5.4 Evaluation of the Kalman filter for tracking the robot’s height . . . . . . 102
5.5 Quantitative evaluation of the Kalman-filtered height estimation . . . . 103
5.6 Elevation mapping during the Rescue Robotics Camp 2006 in Rome . . 104
6.1 Topological map of the disaster area . . . . . . . . . . . . . . . . . . . 113
6.2 A simple graph showing a plan from R1 to G1 (bold edges) . . . . . . . 115
6.3 Exploration scoring trajectories recorded during the finals 2006 . . . . . 123
6.4 Comparing the number of conflicts and travel costs . . . . . . . . . . . 124
6.5 Comparing the locally and globally coordinated exploration . . . . . . . 125
7.1 Hardware for vision processing . . . . . . . . . . . . . . . . . . . . . . 131
7.2 MRF graph online constructed from features . . . . . . . . . . . . . . . 134
7.3 Comparing model complexity of different classifiers . . . . . . . . . . . 135
7.4 Examples from the victim data . . . . . . . . . . . . . . . . . . . . . . 141
8.1 System diagram of the wearable device . . . . . . . . . . . . . . . . . 146
8.2 The wearable emergency-response system . . . . . . . . . . . . . . . . 148
8.3 Online data integration of information from the simulated City of Foligno150
8.4 Comparison between KD-tree clustering and agglomerative clustering . 152
8.5 Successive real-data integration . . . . . . . . . . . . . . . . . . . . . . 156
8.6 The number of found civilians on randomly generated maps . . . . . . 157
8.7 The number of civilian survivors after applying different policies . . . . 159
10.1 Perspective drawing of the 3D LRF mechanics . . . . . . . . . . . . . . 173
10.2 Mechanical drawing 3D LRF: servo mount . . . . . . . . . . . . . . . 173
10.3 Mechanical drawing 3D LRF: U-shaped piece . . . . . . . . . . . . . . 174

List of Figures 193
10.4 Perspective drawings Zerg robot mechanics . . . . . . . . . . . . . . . 174
10.5 Mechanical drawing Zerg robot: inner support frame . . . . . . . . . . 175
10.6 Mechanical drawing Zerg robot: front part . . . . . . . . . . . . . . . . 175
10.7 Mechanical drawing Zerg robot: bottom part . . . . . . . . . . . . . . . 176
10.8 Mechanical drawing Zerg robot: side part . . . . . . . . . . . . . . . . 176
10.9 Mechanical drawing Zerg robot: back part . . . . . . . . . . . . . . . . 177
10.10Mechanical drawing Zerg robot: cover back part . . . . . . . . . . . . . 177
10.11Mechanical drawing Zerg robot: cover front part . . . . . . . . . . . . . 178
10.12Mechanical drawing Zerg robot: assembling cube . . . . . . . . . . . . 178
10.13Perspective drawing of the RFID tag releaser . . . . . . . . . . . . . . 179
10.14Mechanical drawing RFID tag releaser: bottom back part . . . . . . . . 180
10.15Mechanical drawing RFID tag releaser: bottom left part . . . . . . . . . 180
10.16Mechanical drawing RFID tag releaser: bottom right part . . . . . . . . 181
10.17Mechanical drawing RFID tag releaser: slider upper part . . . . . . . . 181
10.18Mechanical drawing RFID tag releaser: slider lower part . . . . . . . . 182
10.19Mechanical drawing RFID tag releaser: servo mount . . . . . . . . . . 182
10.20Mechanical drawing RFID tag releaser: RFID tag container . . . . . . . 183
10.21Schematics of the Zerg robot I . . . . . . . . . . . . . . . . . . . . . . 184
10.22Schematics of the Zerg robot II . . . . . . . . . . . . . . . . . . . . . . 185
10.23Schematics of the Lurker robot I . . . . . . . . . . . . . . . . . . . . . 186
10.24Schematics of the Lurker robot II . . . . . . . . . . . . . . . . . . . . . 187

List of Tables
2.1 Comparison of laser range finders. . . . . . . . . . . . . . . . . . . . . 20
2.2 RFID Communication frequencies . . . . . . . . . . . . . . . . . . . . 29
3.1 Classification accuracy of the slippage detection . . . . . . . . . . . . . 52
3.2 Relative error of visual and conventional wheel odometry . . . . . . . . 54
4.1 Positioning errors form human-robot SLAM . . . . . . . . . . . . . . . 89
6.1 Exploration Results from RoboCup ’06 . . . . . . . . . . . . . . . . . 121
6.2 Victim scoring results from RoboCup ’06 . . . . . . . . . . . . . . . . 122
7.1 Feature combinations of different evidence types . . . . . . . . . . . . 133
7.2 Comparison of the SVM and MRF classifier . . . . . . . . . . . . . . . 137
8.1 Percentage of explored buildings . . . . . . . . . . . . . . . . . . . . . 157
8.2 Percentage of found civilians . . . . . . . . . . . . . . . . . . . . . . . 158
8.3 Number of saved civilians . . . . . . . . . . . . . . . . . . . . . . . . 159
195

Bibliography
[AAAI, 2002] AAAI (2002). In the aftermath of september 11 what roboticists
learned from the search and rescue efforts. http://www.aaai.org/Pressroom/
Releases/release-02-0910.php.
[Administration and Agency, 1995] Administration, U. S. F. and Agency, F. E. M.
(1995). Technical rescue program development manual. Technical report.
[AlienTechnology, 2003] AlienTechnology (2003). Alien technology. http://www.
alientechnology.com/.
[Amft et al., 2004] Amft, O., Lauffer, M., Ossevoort, S., Macaluso, F., Lukowicz, P.,
and Tröster, G. (2004). Design of the QBIC wearable computing platform. In 15th
International Conference on Application-Specific Systems, Architectures and Processors
(ASAP ’04), Galveston, Texas.
[Anguelov et al., 2005] Anguelov, D., Taskar, B., Chatalbashev, V., Koller, D., Gupta,
D., Heitz, G., and Ng, A. (2005). Discriminative learning of markov random fields
for segmentation of 3D range data. In IEEE Computer Society Conference on Vision
and Pattern Recognitionv (CVPR), San Diego, California.
[Arai et al., 2002] Arai, T., Pagello, E., and Parker, L. (2002). Guest editorial advances
in multirobot systems. IEEE Transactions on Robotics and Automation, 18(5):655–
661.
[Bahadori and Iocchi, 2003] Bahadori, S. and Iocchi, L. (2003). Human body detection
in the robocup rescue scenario. In RoboCup 2003: Robot Soccer World Cup VII.
[Bailey, 2002] Bailey, T. (2002). Mobile Robot Localisation and Mapping in Extensive
Outdoor Environments. PhD thesis, Australian Centre for Field Robotics, Australia.
[Balakirsky et al., 2007] Balakirsky, S., Carpin, S., Kleiner, A., Lewis, M., Visser, A.,
Wang, J., and Ziparo, V. (2007). Towards heterogeneous robot teams for disaster
mitigation: Results and performance metrics from robocup rescue. Journal of Field
Robotics, 24(11-12):943–967.
197

198 Bibliography
[Balakirsky et al., 2006] Balakirsky, S., Scrapper, C., Carpin, S., and Lewis., M.
(2006). USARSim: providing a framework for multi-robot performance evaluation.
In Proc. of the Performance Metrics for Intelligent Systems (PerMIS) Workshop.
[Balch and Arkin, 1994] Balch, T. and Arkin, R. C. (1994). Communication in reactive
multiagent robotic systems. Autonomous Robots, 1(1):27–52.
[Bank and Douglas, 1993] Bank, R. E. and Douglas, C. C. (1993). Sparse matrix multiplication
package (SMMP). Advances in Computational Mathematics, 1:127–137.
[Beauregard, 2006] Beauregard, S. (2006). A helmet-mounted pedestrian dead reckoning
system. In Herzog, O., Kenn, H., Lawo, M., Lukowicz, P., and Tröster, G., editors,
3rd International Forum on Applied Wearable Computing (IFAWC’03), pages
79–91. VDE Verlag.
[Beleznai et al., 2004] Beleznai, C., Frühstück, B., and Bischof, H. (2004). Human
detection in groups using a fast mean shift procedure. In Proc. IEEE Int. Conf. on
Image Processing, pages 349–352.
[Bennewitz et al., 2001] Bennewitz, M., Burgard, W., and Thrun, S. (2001). Optimizing
schedules for prioritized path planning of multi-robot systems. In Proc. of the
IEEE Int. Conf. on Robotics & Automation (ICRA), volume 1, pages 271–276.
[Bennewitz et al., 2005] Bennewitz, M., Faber, F., Joho, D., Schreiber, M., and
Behnke, S. (2005). Integrating vision and speech for conversations with multiple
persons. In Proc. of the IEEE/RSJ Int. Conf. on Intelligent Robots & Systems (IROS),
pages 2523–2528, Canada.
[Bentley, 1975] Bentley, J. L. (1975). Multidimensional binary search trees used for
associative searching. Communications of the ACM, 18(9):509–517.
[Bhatt and Glover, 2006] Bhatt, H. and Glover, B. (2006). RFID Essentials. O’Reilly.
[Birk, 2003] Birk, A. (2003). RoboCupRescue - robot league team IUB rescue, Germany.
In RoboCup 2004 (CDROM Proceedings), Team Description Paper, Rescue
Robot League, Bremen, Germany.
[Bock, 1974] Bock, H. (1974). Automatic Classification. Vandenhoeck and Ruprecht,
Göttingen.
[Bohn and Mattern, 2004] Bohn, J. and Mattern, F. (2004). Super-distributed RFID tag
infrastructures. In Proceedings of the 2nd European Symposium on Ambient Intelligence
(EUSAI 2004), number 3295 in Lecture Notes in Computer Science (LNCS),
pages 1–12, Eindhoven, The Netherlands. Springer-Verlag.

Bibliography 199
[Borenstein, 1996] Borenstein, J. (1996). Measurement and correction of systematic
odometry errors in mobile robots. IEEE Transactions on Robotics and Automation,
12(6):869–880.
[Borenstein et al., 1996] Borenstein, J., Everett, H., and Feng, L. (1996). Where am i?
sensors and methods for mobile robot positioning. Technical report, University of
Michigan.
[Bradski, 2000] Bradski, G. (2000). The OpenCV library. Dr. Dobb’s Journal of Software
Tools, 25(11):120, 122–125.
[Bruce, 2000] Bruce, J. (2000). Realtime machine vision perception and prediction.
Master’s thesis, Carnegie Mellon University.
[Burgard et al., 2005] Burgard, W., Moors, M., Stachniss, C., and Schneider, F. (2005).
Coordinated multi-robot exploration. IEEE Transactions on Robotics, 21(3):376–
378.
[Cao et al., 2004] Cao, D., Masoud, O., Boley, D., and Papanikolopoulos, N. (2004).
Online motion classification using support vector machines. In Proc. of the IEEE
Int. Conf. on Robotics & Automation (ICRA), pages 2291–2296.
[Carpin et al., 2006a] Carpin, S., Lewis, M., Wang, J., Balakirsky, S., and Scrapper., C.
(2006a). Bridging the gap between simulation and reality in urban search and rescue.
In Robocup 2006: Robot Soccer World Cup X. Springer, LNAI.
[Carpin et al., 2006b] Carpin, S., Stoyanov, T., Nevatia, Y., Lewis, M., and Wang., J.
(2006b). Quantitative Assessments of USARSim Accuracy. In Proc. of the Performance
Metrics for Intelligent Systems (PerMIS) Workshop.
[Casper et al., 2000] Casper, J., Micire, M., and Murphy, R. (2000). Issues in intelligent
robots for search and rescue. In SPIE Ground Vehicle Technology II, volume 4,
pages 41–46.
[Casper and Murphy, 2003] Casper, J. and Murphy, R. (2003). Human-robot interactions
during the robot-assisted urban search and rescue response at the world trade
center. IEEE Transactions On Systems Man And Cybernetics Part B Cybernetics,
33(3):367–385.
[Co. KG, 2005] Co. KG, V. T. G. . (2005). Vorwerk is presenting the first carpet containing
integrated RFID technology. Press release.
[Corke et al., 2004] Corke, P. I., Strelow, D., and Singh, S. (2004). Omnidirectional visual
odometry for a planetary rover. In Proc. of the IEEE/RSJ Int. Conf. on Intelligent
Robots & Systems (IROS), volume 4, pages 4007–4012.

200 Bibliography
[Cox, 1991] Cox, I. J. (1991). Blanche: Position estimation for an autonomous robot
vehicle. In Iyengar, S. S. and Elfes, A., editors, Autonomous Mobile Robots: Control,
Planning, and Architecture (Vol. 2), pages 285–292. IEEE Computer Society Press,
Los Alamitos, CA.
[CRASAR, 2007] CRASAR (last seen 2007). Return to Katrina. http://crasar.
csee.usf.edu/Research/Projects/Katrina_SGER/index.php.
[Cutler and Davis, 2000] Cutler, R. and Davis, L. S. (2000). Robust real-time periodic
motion detection, analysis, and applications. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 22(8):781–796.
[Dass et al., 2002] Dass, S., Jain, A., and Lu, X. (2002). Face detection and synthesis
using markov random field models. In Proc. of the Int. Conf. on Pattern Recognition
(ICPR), volume 4, pages 201–204, Washington, DC, USA. IEEE Computer Society.
[Davison, 2003] Davison, A. J. (2003). Real-time simultaneous localisation and mapping
with a single camera. In Proc. of the IEEE Int. Conf. on Computer Vision
(ICCV), page 1403, Washington, DC, USA. IEEE Computer Society.
[Dias et al., 2004] Dias, M., Zinck, M., Zlot, R., and Stentz, A. (2004). Robust multirobot
coordination in dynamic environments. In Proc. of the IEEE Int. Conf. on
Robotics & Automation (ICRA), volume 4, pages 3435–3442.
[Dijkstra, 1959] Dijkstra, E. W. (1959). A note on two problems in connexion with
graphs. Numerische Mathematik, 1(1):269–271.
[Dippold, 2006] Dippold, M. (2006). Personal dead reckoning with accelerometers.
In Herzog, O., Kenn, H., Lawo, M., Lukowicz, P., and Tröster, G., editors, 3rd
International Forum on Applied Wearable Computing (IFAWC’03), pages 177–183.
[Dissanayake et al., 2001] Dissanayake, M., Newman, P., Clark, S., Durrant-Whyte,
H., and Csorba, M. (2001). A solution to the simultaneous localization and
map building (SLAM) problem. IEEE Transactions on Robotics and Automation,
17(3):229–241.
[Djugash et al., 2005] Djugash, J., Singh, S., and Corke, P. (2005). Further results with
localization and mapping using range from radio. In Proc. of the Int. Conf. on Field
and Service Robotics (FSR), Pt. Douglas, Australia.
[Dornhege and Kleiner, 2006] Dornhege, C. and Kleiner, A. (2006). Visual odometry
for tracked vehicles. In Proc. of the IEEE Int. Workshop on Safety, Security and
Rescue Robotics (SSRR), Gaithersburg, Maryland, USA.

Bibliography 201
[Dornhege and Kleiner, 2007a] Dornhege, C. and Kleiner, A. (2007a). Behavior maps
for online planning of obstacle negotiation and climbing on rough terrain. In Video
Proc. of the IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS), San
Diego, California.
[Dornhege and Kleiner, 2007b] Dornhege, C. and Kleiner, A. (2007b). Behavior maps
for online planning of obstacle negotiation and climbing on rough terrain. In Proc. of
the IEEE/RSJ Int. Conf. on Intelligent Robots & Systems (IROS), pages 3005–3011,
San Diego, California.
[Durrant-Whyte et al., 1996] Durrant-Whyte, H., Rye, D., and Nebot, E. (1996). Localisation
of automatic guided vehicles. In Robotics Research: The 7th International
Symposium (ISRR’95), pages 613–625. Springer Verlag.
[Elfes, 1989] Elfes, A. (1989). Using occupancy grids for mobile robot perception and
navigation. Computer, 22(6):46–57.
[Eppstein, 2000] Eppstein, D. (2000). Fast hierarchical clustering and other applications
of dynamic closest pairs. J. Exp. Algorithmics, 5:619–628.
[Erdmann and Lozano-Perez, 1987] Erdmann, M. and Lozano-Perez, T. (1987).
multiple moving objects. Algorithmica, 2(2):477–521.
On
[Fachberger et al., 2006] Fachberger, R., Bruckner, G., Reindl, L., and Hauser, R.
(2006). Tagging of metallic objects in harsh environments. In Sensoren und Messsysteme
- 13. ITG-/GMA Fachtagung, University of Freiburg.
[Farinelli et al., 2006] Farinelli, A., Iocchi, L., Nardi, D., and Ziparo, V. (2006). Assignment
of dynamically perceived tasks by token passing in multirobot systems.
Proceedings of the IEEE, 94(7):1271–1288.
[FEMA, 2003] FEMA (2003). National urban search and rescue (USR) response
system - field operations guide. Technical Report US&R-2-FG, U.S. Department
of Homeland Security. http://www.fema.gov/emergency/usr/resources.
shtm.
[Finkenzeller, 2003] Finkenzeller, K. (2003). RFID-Handbook: Fundamentals and
Applications in Contactless Smart Cards and Identification. John Wiley and Sons,
2nd edition.
[Fishkin and Roy, 2003] Fishkin, K. and Roy, S. (2003). Enhancing RFID privacy via
antenna energy analysis. Technical Report IRS-TR-03-012, Intel Research Seattle.
[Foxlin, 2005] Foxlin, E. (2005). Pedestrian tracking with shoe-mounted inertial sensors.
IEEE Comput. Graph. Appl., 25(6):38–46.

202 Bibliography
[Frese, 2006] Frese, U. (2006). Treemap: An O(logn) algorithm for indoor simultaneous
localization and mapping. Autonomous Robots, 21(2):103–122.
[Freund and Schapire, 1996] Freund, Y. and Schapire, R. E. (1996). Experiments
with a new boosting algorithm. In International Conference on Machine Learning
(ICML), pages 148–156.
[Gabaglio, 2003] Gabaglio, V. (2003). GPS/INS integration for pedestrian navigation.
Geodätisch-geophysikalische Arbeiten in der Schweiz, 64:161.
[Gassmann et al., 2003] Gassmann, B., Frommberger, L., Dillmann, R., and Berns, K.
(2003). Real-time 3D map building for local navigation of a walking robot in unstructured
terrain. In Proc. of the IEEE/RSJ Int. Conf. on Intelligent Robots & Systems
(IROS), volume 3, pages 2185–2190.
[Giuliani et al., 1994] Giuliani, D., Omologo, M., and Svaizer, P. (1994). Talker localization
and speech recognition using a microphone array and a cross-powerspectrum
phase analysis. In Proc. of the ICSLP, pages 1243–1246.
[Grewal et al., 2001] Grewal, M., Weill, L. R., and Andrews, A. P. (2001). Global
Positioning Systems, Inertial Navigation, and Integration. John Wiley & Sons.
[Grisetti et al., 2005] Grisetti, G., C., S., and W., B. (2005). Improving grid-based
SLAM with rao-blackwellized particle filters by adaptive proposals and selective
resampling. In Proc. of the IEEE Int. Conf. on Robotics & Automation (ICRA),
pages 667–672, Barcelona, Spain.
[Grisetti et al., 2002] Grisetti, G., Iocchi, L., and Nardi, D. (2002). Global Hough
localization for mobile robots in polygonal environments. In Proc. of the IEEE
Int. Conf. on Robotics & Automation (ICRA), volume 1, pages 353–358.
[Gutmann and Schlegel, 1996] Gutmann, J. and Schlegel, C. (1996). Amos: Comparison
of scan matching approaches for self-localization in indoor environments. In
Proceedings of the 1st Euromicro Workshop on Advanced Mobile Robots, pages 61–
67. IEEE Computer Society Press.
[Gutmann, 2000] Gutmann, J.-S. (2000). Robuste Navigation autonomer mobiler Systeme.
PhD thesis, Albert-Ludwigs-Universit"at Freiburg. ISBN 3-89838-241-9.
[Gutmann et al., 2001] Gutmann, J.-S., Weigel, T., and Nebel, B. (2001). A fast, accurate,
and robust method for self-localization in polygonal environments using laserrange-finders.
Advanced Robotics, 14(8):651–668.
[Hähnel, 2005] Hähnel, D. (2005). Mapping with Mobile Robots. PhD thesis, Universität
Freiburg, Freiburg, Deutschland.

Bibliography 203
[Hähnel et al., 2004] Hähnel, D., Burgard, W., Fox, D., Fishkin, K., and Philipose, M.
(2004). Mapping and localization with RFID technology. In Proc. of the IEEE
Int. Conf. on Robotics & Automation (ICRA), volume 1, pages 1015–1020.
[Hähnel et al., 2003] Hähnel, D., D., Burgard, W., Fox, D., and Thrun, S. (2003). A
highly efficient fastslam algorithm for generating cyclic maps of large-scale environments
from raw laser range measurements. In Proc. of the IEEE/RSJ Int. Conf. on
Intelligent Robots & Systems (IROS), pages 206–211, Las Vegas, USA.
[Han and Bhanu, 2003] Han, J. and Bhanu, B. (2003). Detecting moving humans using
color and infrared video. In Proc. of the IEEE Int. Conf. on Multisensor Fusion and
Integration for Intelligent Systems (MFI), pages 228–233.
[Hancke and Kuhn, 2005] Hancke, G. and Kuhn, M. (2005). An RFID distance bounding
protocol. In First Int. Conf. on Security and Privacy for Emerging Areas in
Communications Networks, pages 67– –73.
[Hatzack and Nebel, 2001] Hatzack, W. and Nebel, B. (2001). Solving the operational
traffic control problem. In Proceedings of the 6th European Conference on Planning
(ECP’01).
[Helmick et al., 2004] Helmick, D., Chang, Y., Roumeliotis, S., Clouse, D., and
Matthies, L. (2004). Path following using visual odometry for a mars rover in highslip
environments. In IEEE Aerospace Conference, volume 2, pages 772–789.
[Hitachi, 2003] Hitachi (2003). Mu Chip Data Sheet. http://www.hitachi-eu.
com/mu/products/mu_chip_data_sheet.pdf.
[Holland, 1975] Holland, J. H. (1975).
University of Michigan Press.
Adaption in Natural and Artificial Systems.
[Honeywell, 2007] Honeywell (2007). Dead-reckoning module DRM 5. Available
on: http://www.ssec.honeywell.com/magnetic/products.html. Referenced
2007.
[IABG mbH, 2007] IABG mbH (2007). Iabg homepage. http://www.iabg.de.
[IEEE, 2002] IEEE (2002). The IEEE standard 802.15.1: Wireless personal area network
standard based on the bluetooth v1.1 foundation specifications.
[Jacoff et al., 2000] Jacoff, A., Messina, E., and Evans, J. (2000). A standard test
course for urban search and rescue robots. In Proc. of the Performance Metrics
for Intelligent Systems (PerMIS) Workshop, pages 253–259.

204 Bibliography
[Jacoff et al., 2001] Jacoff, A., Messina, E., and Evans, J. (2001). Experiences in deploying
test arenas for autonomous mobile robots. In Proc. of the Performance Metrics
for Intelligent Systems (PerMIS) Workshop, Mexico City, Mexico.
[Jacoff et al., 2003] Jacoff, A., Messina, E., Weiss, B., Tadokoro, S., and Nakagawa,
Y. (2003). Test arenas and performance metrics for urban search and rescue robots.
In Proc. of the IEEE/RSJ Int. Conf. on Intelligent Robots & Systems (IROS), pages
27–31, Las Vegas.
[Jaeger and Nebel, 2001] Jaeger, M. and Nebel, B. (2001). Decentralized collision
avoidance, deadlock detection, and deadlock resolution for multiple mobile robots.
In Proc. of the IEEE/RSJ Int. Conf. on Intelligent Robots & Systems (IROS).
[Jennings et al., 1997] Jennings, J., Whelan, G., and Evans, W. (1997). Cooperative
search and rescue with a team of mobile robots. In Proc. of the IEEE Int. Conf. on
Advanced Robotics (ICAR), pages 193–200.
[Judd, 1997] Judd, D. C. T. (1997). A personal dead reckoning module. In Proceedings
of the Institute of Navigation’s GPS Conference, pages 169 – 170.
[Kantor and Singh, 2002] Kantor, G. and Singh, S. (2002). Preliminary results in
range-only localization and mapping. In Proc. of the IEEE Int. Conf. on Robotics
& Automation (ICRA), volume 2, pages 1818–1823, Washington D.C., USA.
[Kantor et al., 2003] Kantor, G., Singh, S., Peterson, R., Rus, D., Das, A., Kumar, V.,
Pereira, G., and Spletzer, J. (2003). Distributed search and rescue with robot and
sensor team. In Proc. of the Int. Conf. on Field and Service Robotics (FSR), pages
327–332. Sage Publications.
[Kehagias et al., 2006] Kehagias, A., Djugash, J., and Singh, S. (2006). Range-only
SLAM with interpolated range data. Technical Report CMU-RI-TR-06-26, Robotics
Institute, Carnegie Mellon University.
[Kenn and Kleiner, 2007] Kenn, H. and Kleiner, A. (2007). Towards the integration of
real-time real-world data in urban search and rescue simulation. In Löffler, J. and
Klann, M., editors, Int. Workshop on Mobile Information Technology for Emergency
Response - MobileResponse 2007, volume 4458 of LNCS, pages 106–115. Springer.
[Kenn and Pfeil, 2006] Kenn, H. and Pfeil, A. (2006). Localizing victims through
sound and probabilistic grid maps in urban search and rescue scenario. In Noda,
I., Jacoff, A., Bredenfeld, A., and Takahashi, Y., editors, Robocup 2005: Robot Soccer
World Cup IX, pages 312–322. Springer.

Bibliography 205
[Kim et al., 2004] Kim, J., Ong, S., Nettleton, E., and Sukkarieh, S. (2004). Decentralised
approach to unmanned aerial vehicle navigation: without the use of GPS
and preloaded map. Journal of Aerospace Engineering, Professional Engineering
Publishing, 218(6):399–416.
[Kitano and Tadokoro, 2001] Kitano, H. and Tadokoro, S. (2001). A grand challenge
for multiagent and intelligent systems. AI Magazine, 22:39–52.
[Kitano et al., 1999] Kitano, H., Tadokoro, S., Noda, I., Matsubara, H., Takahashi, T.,
Shinjou, A., and Shimada, S. (1999). RoboCup Rescue: Search and rescue in largescale
disasters as a domain for autonomous agents research. In IEEE Conf. on Man,
Systems, and Cybernetics(SMC-99).
[Kleiner, 2005] Kleiner, A. (2005). Game AI: The shrinking gap between computer
games and ai systems. Ambient Intelligence: The evolution of technology, communication
and cognition towards the future of human-computer interaction, 6:143–155.
[Kleiner et al., 2006a] Kleiner, A., Behrens, N., and Kenn, H. (2006a). Wearable computing
meets multiagent systems: A real-world interface for the RoboCupRescue
simulation platform. In Jennings, N., Tambe, M., Ishida, T., and Ramchurn, S., editors,
First International Workshop on Agent Technology for Disaster Management at
AAMAS06, pages 116–123, Hakodate, Japan. AAMAS Press.
[Kleiner et al., 2004] Kleiner, A., Brenner, M., Bräuer, T., Dornhege, C., Göbelbecker,
M., Luber, M., Prediger, J., and Stückler, J. (2004). ResQ Freiburg: Team description
and evaluation. In RoboCup 2004 (CDROM Proceedings), Team Description Paper,
Rescue Simulation League, Lisbon, Portugal. (1st place in the competition).
[Kleiner et al., 2005a] Kleiner, A., Brenner, M., Bräuer, T., Dornhege, C., Göbelbecker,
M., Luber, M., Prediger, J., Stückler, J., and Nebel, B. (2005a). Successful
search and rescue in simulated disaster areas. In Bredenfeld, A., Jacoff, A., Noda,
I., and Takahashi, Y., editors, Robocup 2005: Robot Soccer World Cup IX, volume
4020 of Lecture Notes in Computer Science, pages 323–334. Springer.
[Kleiner and Buchheim, 2003] Kleiner, A. and Buchheim, T. (2003). A plugin-based
architecture for simulation in the F2000 league. In Polani, D., Browning, B., Bonarini,
A., and Yoshida, K., editors, RoboCup 2003: Robot Soccer World Cup VII,
volume 3020 of Lecture Notes in Computer Science, pages 434–445. Springer.
[Kleiner and Dornhege, 2007] Kleiner, A. and Dornhege, C. (2007). Real-time localization
and elevation mapping within urban search and rescue scenarios. Journal of
Field Robotics, 24(8–9):723–745.

206 Bibliography
[Kleiner et al., 2006b] Kleiner, A., Dornhege, C., Kümmerle, R., Ruhnke, M., Steder,
B., Nebel, B., Doherty, P., Wzorek, M., Rudol, P., Conte, G., Durante, S., and Lundstrom,
D. (2006b). RoboCupRescue - robot league team RescueRobots Freiburg
(Germany). In RoboCup 2006 (CDROM Proceedings), Team Description Paper,
Rescue Robot League, Bremen, Germany. (1st place in the autonomy competition).
[Kleiner et al., 2007] Kleiner, A., Dornhege, C., and Sun, D. (2007). Mapping disaster
areas jointly: RFID-coordinated slam by humans and robots. In Proc. of the IEEE
Int. Workshop on Safety, Security and Rescue Robotics (SSRR), pages 1–6, Rom,
Italy.
[Kleiner and Kümmerle, 2007] Kleiner, A. and Kümmerle, R. (2007). Genetic MRF
model optimization for real-time victim detection in search and rescue. In Proc. of
the IEEE/RSJ Int. Conf. on Intelligent Robots & Systems (IROS), pages 3025–3030,
San Diego, California.
[Kleiner et al., 2006c] Kleiner, A., Prediger, J., and Nebel, B. (2006c). RFID
technology-based exploration and SLAM for search and rescue. In Proc. of the
IEEE/RSJ Int. Conf. on Intelligent Robots & Systems (IROS), pages 4054–4059, Beijing,
China.
[Kleiner et al., 2005b] Kleiner, A., Steder, B., Dornhege, C., Höfer, D., Meyer-Delius,
D., Prediger, J., Stückler, J., Glogowski, K., Thurner, M., Luber, M., Schnell, M.,
Kümmerle, R., Burk, T., Bräuer, T., and Nebel, B. (2005b). RoboCupRescue - robot
league team RescueRobots Freiburg (Germany). In RoboCup 2005 (CDROM Proceedings),
Team Description Paper, Rescue Robot League, Osaka, Japan. (1st place
in the autonomy competition, 4th place in the teleoperation competition).
[Kleiner and Sun, 2007] Kleiner, A. and Sun, D. (2007). Decentralized SLAM for
pedestrians without direct communication. In Proc. of the IEEE/RSJ Int. Conf. on
Intelligent Robots & Systems (IROS), pages 1461–1466, San Diego, California.
[Kleiner and Ziparo, 2006] Kleiner, A. and Ziparo, V. (2006). RoboCupRescue - simulation
league team RescueRobots Freiburg (Germany). In RoboCup 2006 (CDROM
Proceedings), Team Description Paper, Rescue Simulation League, Bremen, Germany.
(1st place in the competition).
[Kohavi and John, 1997] Kohavi, J. and John, G. H. (1997). Wrappers for feature subset
selection. Artificial Intelligence Journal, 97:273–324.
[Koyanagi et al., 2006] Koyanagi, E., Oba, Y., Kato, J., Sakurada, N., Yoshida, T., and
Hayashibara, Y. (2006). RoboCupRescue 2006 - robot league team toin pelican
(japan). In RoboCup 2006 (CDROM Proceedings), Team Description Paper, Rescue
Robot League, Bremen, Germany.

Bibliography 207
[Krotkov and Hoffman, 1994] Krotkov, E. and Hoffman, R. (1994). Terrain mapping
for a walking planetary rover. IEEE Transactions on Robotics and Automation,
10:728–739.
[Kuhn, 1955] Kuhn, H. W. (1955). The Hungarian method for the assignment problem.
Naval Research Logistics Quaterly, 2:83–97.
[Kurth et al., 2003] Kurth, D., Kantor, G., and Singh, S. (2003). Experimental results in
range-only localization with radio. In Proc. of the IEEE/RSJ Int. Conf. on Intelligent
Robots & Systems (IROS), volume 1, pages 974–979, Las Vegas, USA.
[Kuwata et al., 2006] Kuwata, Y., Takahashi, T., Ito, N., and Takeuchi, I. (2006). Design
of human-in-the-loop agent simulation for disaster simulation systems,. In
Proc. of the Third International Workshop on Synthetic Simulation and Robotics to
Mitigate Earthquake Disaster (SRMED), pages 9–14.
[Ladetto et al., 2001] Ladetto, Q., Gabaglio, V., and Merminod, B. (2001). Combining
gyroscopes, magnetic compass and GPS for pedestrian navigation. In International
Symposium on Kinematic Systems in Geody, Geomatics and Navigation (KIS 2001),
pages 205–212, Banff, Canada.
[Ladetto and Merminod, 2002a] Ladetto, Q. and Merminod, B. (2002a). Digital magnetic
compass and gyroscope integration for pedestrian navigation. In 9th Saint
Petersburg International Conference on Integrated Navigation Systems, Russia.
[Ladetto and Merminod, 2002b] Ladetto, Q. and Merminod, B. (2002b). In step with
INS: Navigation for the blind, tracking emergency crews. GPSWorld, 13(10):30–38.
[Ladetto et al., 2000] Ladetto, Q., Merminod, B., Terrirt, P., and Schutz, Y. (2000). On
foot navigation: When GPS alone is not enough. Journal of Navigation, 53(02):279–
285.
[Lange, 2005] Lange, E. (2005). Sicherheits-innovationen: Banknoten der zukunft.
Die Bank - Zeitschrift für Bankpolitik und Praxis, 9.
[Lawo et al., 2006] Lawo, M., Witt, H., Kenn, H., Nicolai, T., and Leibrand, R. (2006).
A glove for seamless computer interaction - understand the WINSPECT. In Kenn,
H., Glotzbach, U., and Herzog, O., editors, The Smart Glove Workshop, number 33.
[Lechner et al., 2001] Lechner, W., Baumann, S., and Legat, K. (2001). Pedestrian
navigation - bases for the next mass market in mobile positioning. In Proceedings of
Locellus 2001, pages 79–90, Munich, Germany.
[Liao et al., 2007] Liao, L., Patterson, D. J., Fox, D., and Kautz, H. A. (2007). Learning
and inferring transportation routines. Artificial Intelligence, 171(5-6):311–331.

208 Bibliography
[Lienhart and Maydt, 2002] Lienhart, R. and Maydt, J. (2002). An extended set of
haar-like features for rapid object detection. Proc. of the IEEE Int. Conf. on Image
Processing (ICIP), 1:900–903.
[Logitech, 2006] Logitech (2006). Logitech QuickCam Pro 4000. http:
//www.logitech.com/index.cfm/products/details/US/EN,CRID=2204,
CONTENTID=5042.
[Lu and Milios, 1994] Lu, F. and Milios, E. (1994). Robot pose estimation in unknown
environments by matching 2D range scans. In IEEE Computer Society Conference
on Vision and Pattern Recognitionv (CVPR), pages 935–938.
[Lu and Milios, 1997] Lu, F. and Milios, E. (1997). Globally consistent range scan
alignment for environment mapping. Autonomous Robots, 4:333–349.
[Maybeck, 1979] Maybeck, P. S. (1979). Stochastic models, estimation, and control,
volume 141 of Mathematics in Science and Engineering. Academic Press.
[Maybeck, 1990] Maybeck, P. S. (1990). The Kalman filter: An introduction to concepts.
Cox and Wilfong.
[Meier et al., 2005] Meier, D., Stachniss, C., and Burgard, W. (2005). Coordinating
multiple robots during exploration under communication with limited bandwidth. In
Proc. of the European Conference on Mobile Robots (ECMR), pages 26–31, Ancona,
Italy.
[Messina et al., 2005] Messina, E., Jacoff, A., Scholtz, J., Schlenoff, C., Huang, H.,
Lytle, A., and Blitch, J. (2005). Statement of requirements for urban search and
rescue robot performance standards. Technical report, Department of Homeland
Security, Science and Technology Directorate and National Institute of Standards
and Technology.
[MicroSensys, 2007] MicroSensys (2007). Homepage. http://www.microsensys.
de/.
[Milella and Siegwart, 2006] Milella, A. and Siegwart, R. (2006). Stereo-based egomotion
estimation using pixel tracking and iterative closest point. In Proc. of the
IEEE Int. Conf. on Computer Vision Systems (ICVS), page 21, Washington, DC,
USA. IEEE Computer Society.
[Miller et al., 2006] Miller, L., Wilson, P. F., Bryner, N. P., Francis, Guerrieri, J. R.,
Stroup, D. W., and Klein-Berndt, L. (2006). RFID-assisted indoor localization and
communication for first responders. In Proc. of the Int. Symposium on Advanced
Radio Technologies.

Bibliography 209
[Montemerlo et al., 2002] Montemerlo, M., Thrun, S., Koller, D., and Wegbreit, B.
(2002). FastSLAM: a factored solution to the simultaneous localization and mapping
problem. In Eighteenth national conference on Artificial intelligence, pages 593–
598, Menlo Park, CA, USA. American Association for Artificial Intelligence.
[Moravec, 1988] Moravec, H. (1988).
robots. AI Magazine, 9(2):61–74.
Sensor fusion in certainty grids for mobile
[Moravec and Elfes, 1985] Moravec, H. and Elfes, A. (1985). High resolution maps
from wide angle sonar. In Proc. of the IEEE Int. Conf. on Robotics & Automation
(ICRA), pages 116–121.
[Moravec, 1996] Moravec, H. P. (1996). Robot spatial perception by stereoscopic vision
and 3D evidence grids. Technical Report CMU-RI-TR-96-34, Carnegie Mellon
University.
[Murphy, 2000] Murphy, K. (2000). Bayesian map learning in dynamic environments.
In Advances in Neural Information Processing Systems (NIPS). MIT Press.
[Murphy, 2004] Murphy, R. (2004). Human-robot interaction in rescue robotics. Systems,
Man and Cybernetics, Part C: Applications and Reviews, IEEE Transactions
on, 34(2):138–153.
[Murphy et al., 2000] Murphy, R., Casper, J., Micire, M., and Hyams, J. (2000).
Mixed-initiative control of multiple heterogeneous robots for USAR. Technical Report
CRASAR-TR2000-11, Center for Robot Assisted Search & Rescue, University
of South Florida, Tampa, FL.
[Neira and Tard, 2001] Neira, J. and Tard, J. (2001). Data association in stochastic
mapping using the joint compatibility test. IEEE Transactions on Robotics and Automation,
17(6):890–897.
[Nettleton et al., 2003] Nettleton, E., Thrun, S., and Durrant-Whyte, H. (2003). Decentralised
SLAM with low-bandwidth communication for teams of airborne vehicles.
In In Proc. of the Int. Conf. on Field and Service Robotics, Lake Yamanaka, Japan.
[Nevatia et al., 2006] Nevatia, Y., Mahmudi, M., Markov, S., Rathnam, R., Stoyanov,
T., , and Carpin, S. (2006). VIRTUAL-IUB: the 2006 IUB virtual robots team. In
Robocup 2006: Robot Soccer World Cup X, Bremen, Germany.
[Newman et al., 2003] Newman, P. M., Leonard, J. J., and Rikoski, R. R. (2003). Towards
constant-time SLAM on an autonomous underwater vehicle using synthetic
aperture sonar. In Int. Symposium on Robotics Research, Sienna, Italy.

210 Bibliography
[Nguyen et al., 2004] Nguyen, H. M., Takamori, T., Kobayashi, S., Takashima, S., and
Ikeuchi, A. (2004). Development of carbon dioxide sensing system for searching
victims in large scale disasters. In SICE Annual Conference, volume 2, pages 1358–
1361.
[Nicolai et al., 2005] Nicolai, T., Sindt, T., Kenn, H., and Witt, H. (2005). Case study of
wearable computing for aircraft maintenance. In Herzog, O., Lawo, M., Lukowicz,
P., and Randall, J., editors, 2nd International Forum on Applied Wearable Computing
(IFAWC’02), pages 97–110. VDE Verlag.
[Nister et al., 2004] Nister, D., Naroditsky, O., and Bergen, J. (2004). Visual odometry.
In IEEE Computer Society Conference on Vision and Pattern Recognitionv (CVPR),
volume 1, pages 652–659.
[Noda and Meguro, 2006] Noda, I. and Meguro, J. (2006). Data representation for
information sharing and integration among rescue robot and simulation. In International
Joint Conference SICE-ICASE, pages 3449–3454, Busan, Korea.
[Nourbakhsh et al., 2005] Nourbakhsh, I., Lewis, M., Sycara, K., Koes, M., Yong, M.,
and Burion, S. (2005). Human-robot teaming for search and rescue. IEEE Pervasive
Computing, 4(1):72–78.
[Nüchter et al., 2003] Nüchter, A., Surmann, H., Lingemann, K., and Hertzberg, J.
(2003). Semantic scene analysis of scanned 3D indoor environments. In Proceedings
of the 8th International Fall Workshop Vision, Modeling, and Visualization (VMV),
pages 215–222, Munich, Germany. IOS Press.
[Nüssle et al., 2004] Nüssle, T. A., Kleiner, A., and Brenner, M. (2004). Approaching
urban disaster reality: The ResQ firesimulator. In Nardi, D., Riedmiller, M., Sammut,
C., and Santos-Victor, J., editors, RoboCup 2004: Robot Soccer World Cup VIII,
volume 3276 of Lecture Notes in Computer Science, pages 474–482. Springer.
[Oh et al., 2006] Oh, J., Hwang, J., and Smith, S. (2006). Agent technologies for postdisaster
urban planning. In Jennings, N., Tambe, M., Ishida, T., and Ramchurn, S.,
editors, First International Workshop on Agent Technology for Disaster Management
at AAMAS06, pages 24–31, Hakodate, Japan. AAMAS Press.
[Ojeda and Borenstein, 2004] Ojeda, L. and Borenstein, J. (2004). Methods for the
reduction of odometry errors in over-constrained mobile robots. Autonomous Robots,
16:273–286.
[Omologo and Svaizer, 1996] Omologo, M. and Svaizer, P. (1996). Acoustic source
localization in noisy and reverberant environment using csp analysis. In Proc. of

Bibliography 211
the IEEE Int. Conf. on Acoustics, Speech, and Signal Processing (ICASSP), pages
921–924.
[Perkins, 2002] Perkins, C. (2002). IP mobility support for IPv4. RFC.
[Pfaff et al., 2007a] Pfaff, P., Triebel, R., and Burgard, W. (2007a). An efficient extension
to elevation maps for outdoor terrain mapping and loop closing. The International
Journal of Robotics Research - Special Issue FSR, 26(2):217–230.
[Pfaff et al., 2007b] Pfaff, P., Triebel, R., Stachniss, C., Lamon, P., Burgard, W., and
Siegwart, R. (2007b). Towards mapping of cities. In Proc. of the IEEE Int. Conf. on
Robotics & Automation (ICRA), Rome, Italy.
[Pfingsthorn et al., 2006] Pfingsthorn, M., Slamet, B., Visser, A., and Vlassis, N.
(2006). UvA rescue team 2006 RoboCup Rescue - simulation league. In Robocup
2006: Robot Soccer World Cup X, Bremen, Germany.
[Pratt et al., 2006] Pratt, K., Murphy, R., Stover, S., and Griffin, C. (2006). Requirements
for semi-autonomous flight in miniature UAVs for structural inspection. In
Submitted to AUVSI Unmanned Systems North America.
[Quinlan, 2003] Quinlan, J. R. (2003). Induction of decision trees. Machine Learning,
1(1):81–106.
[Ramos et al., 2007] Ramos, F. T., Nieto, J., and Durrant-Whyte, H. (2007). Recognising
and modelling landmarks to close loops in outdoor SLAM. In Proc. of the IEEE
Int. Conf. on Robotics & Automation (ICRA), pages 2036–2041.
[Reece and Roberts, 2005] Reece, S. and Roberts, S. (2005). Robust, low-bandwidth,
multi-vehicle mapping. In Eighth IEEE Int. Conf. on Information Fusion, Philadelphia,
Penn.
[Reindl et al., 2001] Reindl, L. M., Pohl, A., Scholl, G., and Weigel, R. (2001). Sawbased
radio sensor systems. IEEE Sensors Journal, 1:69–78.
[Research, 2007] Research, C. T. (2007). www.ctr.at.
[ResQ Freiburg, 2004] ResQ Freiburg (2004). Source code. Available on: http://
gkiweb.informatik.uni-freiburg.de/~rescue/sim04/source/resq.tgz.
release September, 2004.
[Retscher and Kealy, 2006] Retscher, G. and Kealy, A. (2006). Ubiquitous positioning
technologies for modern intelligent navigation systems. The Journal of Navigation,
59:91–103.

212 Bibliography
[Russell and Norvig, 2003] Russell, S. J. and Norvig, P. (2003). Artificial Intelligence:
A Modern Approach. Pearson Education.
[Schurr et al., 2005] Schurr, N., Marecki, J., Scerri, P., Lewi, J. P., and Tambe, M.
(2005). Programming Multiagent Systems, chapter The DEFACTO System: Coordinating
Human-Agent Teams for the Future of Disaster Response, page 296.
Springer.
[Seidel and Rapport, 1992] Seidel, S. Y. and Rapport, T. S. (1992). 914 MHz path
loss prediction model for indoor wireless communications in multi-floored buildings.
IEEE Trans. on Antennas and Propagation, 40(2):207–217.
[Sheh, 2006] Sheh, R. (2006). The Redback: A low-cost advanced mobility robot for
education and research. In Proc. of the IEEE Int. Workshop on Safty, Security and
Rescue Robotics (SSRR).
[Shimony, 1994] Shimony, S. E. (1994). Finding maps for belief networks is np-hard.
Artificial Intelligence, 68:399–410.
[Sidenbladh, 2004] Sidenbladh, H. (2004). Detecting human motion with support vector
machines. In Proc. of the Int. Conf. on Pattern Recognition (ICPR), volume 2,
pages 188–191.
[Siebra and Tate, 2005] Siebra, C. and Tate, A. (2005). Integrating collaboration and
activity-oriented planning for coalition operations support. In In Proceedings of the
9th International Symposium on RoboCup, Osaka, Japan.
[Smith et al., 1988] Smith, R., Self, M., and Cheeseman, P. (1988). Estimating uncertain
spatial relationships in robotics. Autonomous Robot Vehicles, 1:167–193.
[Smith and Cheeseman, 1986] Smith, R. C. and Cheeseman, P. (1986). On the representation
and estimation of spatial uncertainty. International Journal of Robotics
Research, 5(4):56–68.
[Snyder, 1987] Snyder, J. P. (1987). Map Projections - A Working Manual. U.S. Geological
Survey Professional Paper 1395. United States Government Printing Office,
Washington, D.C.
[Stachniss et al., 2004] Stachniss, C., Haehnel, D., and Burgard, W. (2004). Exploration
with active loop-closing for FastSLAM. In Proc. of the IEEE/RSJ Int. Conf.
on Intelligent Robots & Systems (IROS), pages 1505–1510, Sendai, Japan.
[Starner, 2001a] Starner, T. (2001a). The challenges of wearable computing: Part 1.
IEEE Micro, 21(4):44–52.

Bibliography 213
[Starner, 2001b] Starner, T. (2001b). The challenges of wearable computing: Part 2.
IEEE Micro, 21(4):54–67.
[Surmann et al., 2003] Surmann, H., Nüchter, A., and Hertzberg, J. (2003). An autonomous
mobile robot with a 3D laser range finder for 3D exploration and digitalization
of indoor environments. Journal Robotics and Autonomous Systems, 45(3–
4):181–198.
[Surmann et al., 2004] Surmann, H., Worst, R., Hennig, M., Lingemann, K., Nüchter,
A., Pervoelz, K., Tiruchinapalli, K. R., Christaller, T., and Hertzberg, J. (2004).
RoboCupRescue - robot league team KURT3D, Germany, team description paper,
rescue robot league competition. In RoboCup 2004: Robot Soccer World Cup VIII,
Portugal, July.
[Sutton and Barto, 1998] Sutton, R. and Barto, A. G. (1998). Reinforcement Learning:
An Introduction (Adaptive Computation and Machine Learning). The MIT Press.
[Svaizer et al., 1997] Svaizer, P., Matassoni, M., and Omologo, M. (1997). Acoustic
source location in a three-dimensional space using crosspower spectrum phase. In
Proc. of the IEEE Int. Conf. on Acoustics, Speech, and Signal Processing (ICASSP),
page 231, Washington, DC, USA. IEEE Computer Society.
[Svennebring and Koenig, 2004] Svennebring, J. and Koenig, S. (2004). Building
terrain-covering ant robots: A feasibility study. Auton. Robots, 16(3):313–332.
[Tadokoro, 2005] Tadokoro, S. (2005). Special project on development of advanced
robots for disaster response (DDT project). In IEEE Workshop on Advance Robotics
and its Social Impacts (ARSO’05).
[Tadokoro et al., 2000] Tadokoro, S., Kitano, H., Takahashi, T., Noda, I., Matsubara,
H., Shinjou, A., Koto, T., Takeuchi, I., Takahashi, H., Matsuno, F., Hatayama, M.,
Nobe, J., and Shimada, S. (2000). The RoboCup-Rescue project: A robotic approach
to the disaster mitigation problem. In Proc. of the IEEE Int. Conf. on Robotics &
Automation (ICRA).
[Tadokoro et al., 2003] Tadokoro, S., Matsuno, F., Onosato, M., and Asama, H. (2003).
Japan national special project for earthquake disaster mitigation in urban areas.
In First International Workshop on Synthetic Simulation and Robotics to Mitigate
Earthquake Disaster.
[Takahashi, 2006] Takahashi, T. (2006). Requirements to agent-based disaster simulations
from local government usage. In Jennings, N., Tambe, M., Ishida, T., and
Ramchurn, S., editors, First International Workshop on Agent Technology for Disaster
Management at AAMAS06, pages 78–84, Hakodate, Japan. AAMAS Press.

214 Bibliography
[Takashi and Hiroshi, 2003] Takashi, N. and Hiroshi, U. (2003). Detection of carbon
dioxide in urban search and rescue as an application for miniaturized mass spectrometerst.
Journal of the Mass Spectrometry Society of Japan, 51(1):264–266.
[Taskar et al., 2004] Taskar, B., Chatalbashev, V., and Koller, D. (2004). Learning associative
markov networks. In Proc. of the Int. Conf. on Machine Learning (ICML).
[Taskar et al., 2003] Taskar, B., Guestrin, C., and Koller, D. (2003). Max-margin
markov networks. In Advances in Neural Information Processing Systems (NIPS),
Vancouver, Canada.
[Thrun et al., 2005] Thrun, S., Burgard, W., and Fox, D. (2005).
Robotics. MIT Press.
Probabilistic
[Thrun et al., 2004] Thrun, S., Liu, Y., Koller, D., Ng, A., Ghahramani, Z., and
Durrant-Whyte, H. (2004). Simultaneous mapping and localization with sparse extended
information filters. International Journal of Robotics Research, 23(7–8):693–
716.
[Thrun et al., 2006] Thrun, S., Montemerlo, M., Dahlkamp, H., Stavens, D., Aron, A.,
Diebel, J., Fong, P., Gale, J., Halpenny, M., Hoffmann, G., Lau, K., Oakley, C.,
Palatucci, M., Pratt, V., Stang, P., Strohband, S., Dupont, C., Jendrossek, L.-E., Koelen,
C., Markey, C., Rummel, C., van Niekerk, J., Jensen, E., Alessandrini, P., Bradski,
G., Davies, B., Ettinger, S., Kaehler, A., Nefian, A., and Mahoney, P. (2006).
Stanley, the robot that won the DARPA grand challenge. Journal of Field Robotics,
23(9):655–656.
[Tomasi and Kanade, 1991] Tomasi, C. and Kanade, T. (1991). Detection and tracking
of point features. Technical Report CMU-CS-91-132, Carnegie Mellon University.
[TopoGrafix, 2004] TopoGrafix (2004). GPX - the GPS exchange format. Available
on: http://www.topografix.com/gpx.asp. release August, 9th 2004.
[Tsai, 1986] Tsai, R. Y. (1986). An efficient and accurate camera calibration technique
for 3D machine vision. In IEEE Computer Society Conference on Vision and Pattern
Recognitionv (CVPR), pages 364–374.
[U and Reed., 2006] U, V. R. and Reed., N. E. (2006). Enhancing agent capabilities in
a large rescue simulation system. In Jennings, N., Tambe, M., Ishida, T., and Ramchurn,
S., editors, First International Workshop on Agent Technology for Disaster
Management at AAMAS06, pages 71–77, Hakodate, Japan. AAMAS Press.
[Vapnik, 1995] Vapnik, V. N. (1995). The nature of statistical learning theory.
Springer-Verlag New York, Inc., New York, NY, USA.

Bibliography 215
[Vectronix, 2007] Vectronix (2007). Personal dead reckoning module. Available on:
http://www.vectronix.ch/. Referenced 2007.
[Viola and Jones, 2001] Viola, P. and Jones, M. J. (2001). Rapid object detection using
a boosted cascade of simple features. In IEEE Computer Society Conference on
Vision and Pattern Recognitionv (CVPR).
[Viola et al., 2003] Viola, P., Jones, M. J., and Snow, D. (2003). Detecting pedestrians
using patterns of motion and appearance. In IEEE International Conference on
Computer Vision, volume 2.
[Wang and Lewis, 2007] Wang, J. and Lewis, M. (2007). Human control of cooperating
robot teams. In Human-Robot Interaction Conference (HRI).
[Weigel et al., 2002] Weigel, T., Gutmann, J.-S., Dietl, M., Kleiner, A., and Nebel,
B. (2002). CS Freiburg: Coordinating robots for successful soccer playing. IEEE
Transactions on Robotics and Automation, 18(5):685–699.
[Wickler et al., 2006] Wickler, G., Tate, A., and Potter, S. (2006). Using the constraint model as a shared representation of intentions for emergency response
coordination. In Jennings, N., Tambe, M., Ishida, T., and Ramchurn, S., editors,
First International Workshop on Agent Technology for Disaster Management at AA-
MAS06, pages 2–9, Hakodate, Japan. AAMAS Press.
[Witt et al., 2006] Witt, H., Nicolai, T., and Kenn, H. (2006). Designing a wearable
user interface for hands-free interaction and maintenance applications. In Fourth
Annual IEEE International Conference on Pervasive Computer and Communication
(PerCom).
[Wolf et al., 2005] Wolf, D., Sukhatme, G., Fox, D., and Burgard, W. (2005). Autonomous
terrain mapping and classification using hidden markov models. In
Proc. of the IEEE Int. Conf. on Robotics & Automation (ICRA), pages 2026–2031.
[Wren et al., 1997] Wren, C. R., Azarbayejani, A., Darrell, T., and Pentland, A. (1997).
Pfinder: Real-time tracking of the human body. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 19(7):780–785.
[Wu and Yu, 2006] Wu, Y. and Yu, T. (2006). A field model for human detection
and tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence,
28(5):753–765.
[Yamauchi, 1997] Yamauchi, B. (1997). A frontier-based approach for autonomous
exploration. In IEEE International Symposium on Computational Intelligence in
Robotics and Automation (CIRA ’97).

216 Bibliography
[Ye and Borenstein, 2003] Ye, C. and Borenstein, J. (2003). A new terrain mapping
method for mobile robots obstacle negotiation. In Gerhart, G. R., Shoemaker, C. M.,
and Gage, D. W., editors, Unmanned Ground Vehicle Technology V., volume 5083,
pages 52–62.
[Zalud, 2004] Zalud, L. (2004). Orpheus - universal reconnaissance teleoperated robot.
In RoboCup 2004: Robot Soccer World Cup VIII, pages 491–498.
[Zhu et al., 2006] Zhu, Z., Oskiper, T., Naroditsky, O., Samarasekera, S., Sawhney,
H. S., and Kumar, R. (2006). An improved stereo-based visual odometry system.
In Proc. of the IEEE Int. Workshop on Safty, Security and Rescue Robotics (SSRR),
Gaithersburg, Maryland, USA.
[Ziparo and Iocchi, 2006] Ziparo, V. and Iocchi, L. (2006). Petri net plans. In Proc. of
ATPN/ACSD Fourth International Workshop on Modelling of Objects, Components,
and Agents.
[Ziparo et al., 2007a] Ziparo, V., Kleiner, A., Marchetti, L., Farinelli, A., and Nardi, D.
(2007a). Cooperative exploration for USAR robots with indirect communication. In
Proc.of 6th IFAC Symposium on Intelligent Autonomous Vehicles (IAV’07).
[Ziparo et al., 2007b] Ziparo, V., Kleiner, A., Nebel, B., and Nardi, D. (2007b). RFIDbased
exploration for large robot teams. In Proc. of the IEEE Int. Conf. on Robotics
& Automation (ICRA), pages 4606–4613.
[Zlot et al., 2002] Zlot, R., Stentz, A., Dias, M., and Thayer, S. (2002). Multi-robot
exploration controlled by a market economy. In Proc. of the IEEE Int. Conf. on
Robotics & Automation (ICRA), pages 3016–3023.